How AI search works

• A lot of ChatGPT's answers come from training data. For "who should I hire" questions, it runs web searches.
• For a local-business question, ChatGPT looks up business listings and runs an ordinary web search in one call. Among 27,405 local answers we analyzed, 95% included a card of businesses.
• Across 116 local answers we captured, the search returned a median of 12 web pages per answer, and ChatGPT cited 2 of them at the median.
• Directories and Yelp carry considerable weight in local answers. Angi appeared as a source in 16.5% of 41,160 local answers, and the star rating on the card agreed with Yelp's for more businesses than it agreed with Google's. Yelp's share of those star-rating matches never exceeded half.
• There's no rank in an AI answer. Across 188,520 pairs where we asked ChatGPT the same question twice, minutes apart, the identical set of businesses appeared 3.6% of the time.
• The order matters. First, let AI crawlers read your site. Then keep your business listing accurate and get listed on the directories and review sites AI answers cite. Last, measure how often you're named and report it with a margin of error.

What this guide is built on

Every figure in the guide comes from a dataset in the table or a primary source linked where the figure appears. We keep the names of businesses in our data private.

How to read the evidence

• Measured figures come from our own data. We give the sample size and dates with each one.
• Verified claims have been checked against a primary source, such as OpenAI or Google documentation. We link the source where the claim appears.
• Everything else is our interpretation of the evidence, and it isn't a tested result.

1. Part 01You ask a question

   Who’s the best [service] in [City]?

2. Part 02The engine goes looking

   Business listings

   Web search

3. Part 03It picks what to use

   12 returned, 2 cited

4. Part 04It writes the answer

   Business A★ 4.8

   Business B★ 4.6

ChatGPT looks up business listings and searches the web at once, brings back 12 pages, and cites 2 of them in the answer.

Part 1 You ask a question

Where an AI answer gets its information

An AI answer can draw on information absorbed during training, a search index assembled ahead of time, a live page visit while you wait, or data licensed from a company. Small local businesses have little presence in training data, so the search routes carry most of the load for them.

What the model already knows

The model takes in your question as tokens, which OpenAI describes as "a character, part of a word, a whole word, or punctuation," with each token roughly four characters in English. What the model learns during training is held in its weights, the "large sets of numbers" that change as it learns. A 2023 study found that models answer a factual question better when more training documents relate to it. A small local business with few online mentions is therefore thin in the model's memory.

ChatGPT's current models reason before they answer. OpenAI's developer docs say the models "use internal reasoning tokens before producing a response" and that OpenAI keeps the raw reasoning hidden. In 116 ChatGPT conversations we captured about local questions, the only visible sign of that step was a one-line status such as "Searching 7 websites."

The underlying model changes often. OpenAI's model release notes list 15 dated entries from January through August 2026, covering new models, updates, and retirements. What a model knows about your business from training shifts slowly, and no one outside the company can measure it.

How new information gets in

OpenAI's web search docs say its tool can fetch "live content" or use only "cached/indexed results" when searching, and Search Engine Land reported that Yelp struck a July 2026 licensing deal with OpenAI covering reviews, ratings, photos, and business info. ChatGPT decides when to search, and OpenAI says it "may search the web automatically when your question would benefit from current information." For local questions, we saw it search every time its search call was visible to us, in 100 out of 100 cases. We also put 92 general AI-search questions to ChatGPT in three sittings on September 24 and 25, 2026, and it handled 79 the same way each time, using the web for 58 in all three sittings and for 21 in none. Those 21 included 8 of the 10 definition questions, while all 8 questions about AI visibility tools and both questions about hiring an agency used the web in every sitting.

Switching web search off doesn't turn a local answer into pure memory. In a test of local questions through a commercial scraper with its web search setting off, ChatGPT's card of businesses appeared in 95.5% of 3,303 answers.

ChatGPT's memory concerns the person asking. OpenAI says it includes preferences and details from that person's earlier chats, saved memories, and connected apps such as Gmail. That memory can shape what one customer sees, but it isn't a record of your business that you can update.

What is query fan-out, and what does ChatGPT actually search for?

Query fan-out is when an engine breaks one question into multiple searches it writes itself. Google says AI Overviews and AI Mode can do this by "issuing multiple related searches across subtopics and data sources." OpenAI says ChatGPT "typically rewrites your query into one or more targeted queries" and may send "additional, more specific queries" after reviewing the first results. In raw conversation data we captured from a logged-in browser, search appears as a tool the model calls with words it chooses, so your business has to match a query the customer never wrote.

Local fan-out was small across all 100 logged-in captures with a readable search call from Sept. 8 and 10, 2026. Before writing any answer text, ChatGPT made a single search call pairing a business-listings index lookup with an ordinary web search running in parallel. The lookup held a place and one to four category phrases, while the web side held one query in 92 calls and two in the other eight. Every answer also made an earlier call that repeated the customer's question word for word. Our post, the six steps behind one answer, traces the rest of that data.

Both parts of the search call added words the customer hadn't typed. "Top rated" appeared in 12 of the 100 lookups and "licensed" in two, although neither phrase appeared in any of the 100 questions. "Best" appeared in 38 lookups, and in every case it came from the customer's question. On the web side, 70 of the 111 queries used "review," a word absent from all 100 questions.

best plumber in Austinbest plumber in Austin

Business listings

Austin, TX, USA

best plumbers

top rated plumbers

emergency plumbers

Web search

best plumbers Austin TX reviews 2026

For a search such as “best plumber in Austin,” ChatGPT sends two searches in one call, adding the words in blue.

The commercial scraper used for our census gives us a thinner view of search wording. Search queries survived in its records for 877 of 44,988 answers, yielding 1,347 queries in total. Of those queries, 91.7% included a city, 96.7% included a trade, and 31.9% included "review." The scraper and logged-in browser differ in both the questions they received and how they captured searches, so we can't compare that 31.9% directly with the logged-in browser's 70 of 111 queries.

The added words also changed with the question. On Sept. 24, 2026, we asked 92 questions about AI search itself once each on a Plus account in one sitting. Fifty-nine answers made a search call, producing 147 queries. Among the 147 queries, 49 used site:, a command restricting a search to one website. Of those 49, 42 targeted a company named in the question, while 40 of the 147 queries added "official." None of the questions contained site: or "official."

The engine added "2026" to 21 queries across 19 questions that hadn't named a year. We asked the same questions in two further sittings, and the search topic held steady while the wording changed. Among the 56 to 61 questions with a search in both sittings of a pair, only three to five repeated any one query word for word, while roughly two-thirds of one sitting's query words recurred in the other. site: or "official" still appeared in 56% to 62% of queries per sitting.

Since Sept. 23, the lookup line has been harder to see. The saved search call was empty in that day's captures, although web queries remained readable through a separate field, and the call text was still empty in captures made the following day. This looks like a change on ChatGPT's end. We don't yet know whether it stopped sending the lookup or only stopped storing it.

How the way someone asks changes who ChatGPT names

In our census of local answers, changing only the wording of a question changed how many businesses ChatGPT named. The comparison used 3,915 answers for each phrasing. "List three" averaged 3.4 businesses, the plain "best [trade] in [city]" averaged 5.8, and a paragraph-length request that asked for a locally owned firm and excluded chains, franchises, and directory sites averaged 10.6. A directory was named or cited in 94.9% of answers to the plain question, compared with 46.1% for the locally owned version.

Longer questions change the shape of an answer. Google says the average AI Mode search is "triple the length of a traditional Search query." In our repeat study, a card appeared in 78% to 85% of answers to paragraph-long questions that described a job and named a particular service, compared with 90% to 98% for short questions like "best plumber in Austin." The pattern held at all five collection points, with 400 answers per question type at each point.

Two people can ask the same thing and get different businesses in the answer. OpenAI says ChatGPT may use "an approximate location based on your IP address" for local results, and saved memories can feed into how it rewrites a search. Customers ask follow-up questions as well, though we've found no public measurement of how often they do.

We've found no published, verified count of people using AI to pick a plumber, dentist, or lawyer. Adoption surveys we've looked at, including G2's, cover business software buyers. We don't apply those numbers to local services.

Part 2 The engine goes looking

How ChatGPT, Gemini, Claude, and Perplexity find businesses

In our August 2026 cross-engine run, only ChatGPT returned a structured card of businesses for the same 10 local prompts, while the other engines answered in prose with citations. The run included 2,560 ChatGPT answers and 1,200 answers per other surface. We read Gemini's consumer app through a scraper, so an on-screen card may have been lost during capture. Claude and Perplexity went through their APIs, which are developer versions and may differ from the consumer apps.

The engines search in different places, and OpenAI says ChatGPT search "sometimes partners with other search providers" on a page linking to Microsoft's and Shopify's privacy policies for those partners. We haven't measured which index ChatGPT's local searches use. Gemini's API returned 38,307 citations, each one a Google grounding link, so Gemini retrieves through Google. Perplexity says its own search system cuts pages into passages and reranks them.

Brave powers Claude's search, according to our ten-week Brave study. Across 1,481 matched queries from June to August 2026, 87.8% of pages retrieved by Claude's search tool ranked in Brave's top 10 for the query, compared with 45.6% in Google's top 10. Brave rank also tracked whether Claude cited a business. Across 1,408 observations on 52 client accounts, Claude cited businesses in 92.0% of answers when they ranked 1 to 3 in Brave, 69.8% at ranks 4 to 10, 38.4% at ranks 11 to 20, and 11.9% outside the top 20.

Despite their different search sources, all five engine surfaces cited Angi, BBB, HomeAdvisor, and Expertise.com in the August run. Yelp citations varied across engines, with Claude citing yelp.com in 33.2% of answers and Perplexity in 29.3%, while ChatGPT cited it in none of its 2,560 answers. Yelp still reached ChatGPT's answers through ChatGPT's cards. Yelp's rating logo appeared on 27.9% of 215 card rows in our logged-in browser captures, consistent with the licensing deal covered in Part 1.

Apple has its own crawler. Apple says Applebot's data feeds search in Spotlight, Siri, and Safari, and "may be used to provide additional context and up-to-date content" when Apple's AI models write answers. Applebot-Extended is a robots.txt control over whether a site's pages can be used to train Apple's models, and Apple says it "does not crawl webpages." We haven't measured answers from Siri, Apple Intelligence, or Copilot.

We treat each engine as a separate channel, with overlapping work at the directories they all cite. A single score blended across engines hides which engine your business is absent from.

Each share in the table is the fraction of answers citing that site in the August 2026 cross-engine run, which used the same 10 prompts across surfaces. For ChatGPT, the share comes from the scraper's list of cited sources. That list leaves out the Yelp ratings on ChatGPT's card.

Do Google's AI Overviews show up on local searches?

AI Overviews rarely show up on a plain local search, and a 2026 study of medical practices across five US metros found them on under 2% of "find a provider" searches, with map results staying first. When searches added a condition, urgency, or an insurance angle, the rate climbed into the 22.5% to 42.2% range. Across 22.4 million local searches, Ahrefs put it at 7.9%. Google says Overviews appear only when its systems deem them "additive to classic Search," and "as such, often don't trigger."

Google describes AI Overviews as a quick way to get the core of a question, with links for going deeper, while AI Mode is meant for questions needing "further exploration, reasoning, or complex comparisons." Google says both features rest on its core ranking systems, draw pages from the Google index, and may split a question into several searches. A page can be linked only if Google has it in its index and it qualifies for the preview shown with a regular search result. We haven't yet measured AI Mode on local service questions.

When an Overview appears, fewer people click. In Pew's March 2025 tracking, a panel of 900 US adults clicked a regular result on 8% of Google visits with an AI summary, compared with 15% of visits without one, and clicked a link inside the summary on 1% of summary visits. Pew measured every kind of Google search, while plain local searches seldom trigger an Overview. Each site now has a Search Console report showing how often its pages appeared in AI Overviews and AI Mode, which Google had rolled out to all websites by Aug. 31, 2026.

Where does ChatGPT get its information about local businesses?

A local ChatGPT answer contains a listings card and web pages, and ChatGPT fetches both at the same time. In every one of our 100 logged-in captures with a readable search call, the listings lookup and web search shared the same call. Only the lookup took the form of a places search, and the card appeared in 97 of the 100 answers. The data doesn't connect the lookup to the card, so our view that the lookup fills the card remains untested.

The lookup named a real place in 98 of the 100 captures, such as "Austin, TX, USA" for an Austin question. The other two contained the literal string "user," and we don't yet know what it means. For "near me" questions, OpenAI says ChatGPT can estimate a user's location from their IP address and that ChatGPT "may share your location with trusted third-party providers that help provide more local information, such as nearby business listings." Our view is that a near-me search begins at the customer's location.

Each business arrives as a 57-field record, and in a September 23, 2026 check, we found the same 57 fields in each of 173 records from 16 captures. The card shows eight of those fields, including the business name, photo, category, star rating and its rating logo, position on the map, and open-now status. Tap a business and a side panel opens with more of those fields, such as its address, phone number, and hours. The record also includes whether the listing is claimed and review highlights.

Placeholder Plumbing Co.

★ 4.8

Plumber

Open

Its record57 fields

• id
• entity_key
• entity_lookup_id
• provider
• provider_url
• provider_logo_url
• provider_logo_dark_url
• provider_logo_url_short, drawn on the card
• provider_logo_dark_url_short
• name, drawn on the card
• latitude, drawn on the card
• longitude, drawn on the card
• distance_meters
• ranking_score
• rating_scale
• address
• description
• description_cite
• image_url, drawn on the card
• image_urls
• provider_images
• reconciled_thumbnails
• selected_images
• rating, drawn on the card
• review_count
• reviews
• review_highlights
• price
• price_str
• categories, drawn on the card
• website_url
• hours
• is_open, drawn on the card
• next_open_hour
• phone
• is_closed_permanently
• is_closed_temporarily
• rank
• country_code
• country_name
• city
• state
• zipcode
• location
• tags
• attributes
• is_claimed
• date_opened
• date_closed
• special_hours
• menu
• yelp_menu_url
• popularity_score
• reservation_providers
• service_providers
• enriched_description
• from_cache

ChatGPT gets 57 fields for each business. The card shows 8 of them.

Those rows have the shape of a listing, with a name, address, phone number, and category. Business identity follows Google's listing data, though the data doesn't name the provider behind that layer. A logo field on each row identifies where its star rating came from. Part 3 looks at how often that source is Yelp.

In our census of 44,988 local answers, we counted 276,628 card rows. Of those rows, 62% had an address in the metro's principal city, 95% had an address somewhere in the metro, 56% had a category matching the requested trade, and 64% had a website attached. That leaves 36% of card rows without a website. The listing gets a business onto the card, so first check whether its address is in the city the customer named and whether its primary category matches the requested trade, though we haven't tested either condition.

The card doesn't contain every business ChatGPT names. In 70.4% of answers with a card, ChatGPT named at least one business absent from that answer's card. Part 3 examines the web pages cited alongside those extra names.

Can AI crawlers read my site?

About one in four local business sites is unreadable to at least one AI crawler.

AI companies send three kinds of visitors to websites. Training crawlers, such as OpenAI's GPTBot and Anthropic's ClaudeBot, collect pages for possible use in training later models. Search crawlers, such as OAI-SearchBot and Claude-SearchBot, build the index an assistant searches, while user-triggered fetchers, such as ChatGPT-User, Claude-User, and Perplexity-User, load a page when someone is waiting for an answer. We maintain a fuller list of AI crawlers.

OpenAI handles its crawlers separately and says a site that blocks OAI-SearchBot "will not be shown in ChatGPT search answers," though it "can still appear as navigational links." Blocking GPTBot tells OpenAI you don't want your content used for training. A person starts a user-triggered fetch, so OpenAI says robots.txt rules "may not apply" to those visits, while Perplexity says its fetcher "generally ignores" those rules. Google's AI features use regular Googlebot, so there's no separate crawler to allow or block for AI Overviews.

In an access census collected August 20, 2026, we checked 15,834 live sites belonging to independent local businesses across 40 US metros and found that 23.77% were unreadable to at least one of the 11 AI crawlers we tested. The figure was 27.06% when weighted to match the full sample's makeup. Sites failed because of a robots.txt rule, a block or error from the host or CDN (the network layer in front of many sites), or raw homepage HTML that didn't contain both what the business does and a contact detail or location. Among the 13,760 sites that loaded for a regular browser, roughly 9% refused a request identifying itself as an AI crawler, with the rate ranging from 8.0% to 11.0% depending on how we counted rate-limit responses.

OpenAI says "any public website can appear in ChatGPT search." Its guidance is to allow OAI-SearchBot in robots.txt and confirm that your host or CDN allows OpenAI's published IP addresses. OpenAI says robots.txt changes take roughly 24 hours to reach ChatGPT search. Server logs record some crawler visits, but a quiet log doesn't prove your site wasn't used, because many answers draw on copies an engine saved earlier.

Do I need an llms.txt file?

For a local business, an llms.txt file is optional. It lives at yoursite.com/llms.txt and is intended to give AI crawlers a plain-text summary of a site's key content. Google says its Search makes no use of llms.txt files and that creating one "will neither harm nor help your site's visibility or rankings in Google Search." Google takes the same position on Markdown versions of pages and other AI-only files.

We measured which AI bots fetch the file across 6,455 WordPress sites with our llms.txt plugin that had opted in to sharing bot data. In the study's last 30 days, ending June 14, 2026, 626 of those sites (9.7%) received a recognized AI bot request for /llms.txt. OpenAI's crawlers led, with OAI-SearchBot requesting the file on 8.5% of the 5,567 sites it crawled and GPTBot on 8.4% of the 4,434 sites it crawled. Fetchers working while a person waits almost never requested the file, with ChatGPT-User doing so on 0.4% of the 5,634 sites it reached and Perplexity-User on none of the 2,895 sites it reached, while most bots that fetched the file came once and didn't return.

The study counted fetches only, so it didn't test whether having the file changes whether a business gets named or cited. We recorded zero Google-Extended fetches, as expected, because Google says Google-Extended "doesn't have a separate HTTP request user agent string" and is a robots.txt switch that decides whether Gemini may use a site's content for training and to inform its answers. The file does no harm and probably won't change a local AI answer, since the card showing business details uses listing data and the written answer cites directories and business sites. A homepage crawlers can read and an accurate listing come first.

Does ChatGPT read my page at all?

Roughly one time in four, no AI assistant reads the site while a customer waits for an answer. The verifyagents.org study looked at 22 small or midsize business sites and found that, for 26% of 128 clicks from AI answers that carried a tracking tag, no AI assistant had fetched the site during the preceding seven days. The study measured the missing fetch alone and offers an earlier index as its best guess for where the answer came from. Google calls the practice of linking an answer to retrieved pages "grounding," and OpenAI's documentation says its search tool may use "cached/indexed results" as well as live content.

Even when an engine reads a page, it may use only part of it. Perplexity says it breaks pages into standalone passages and ranks each passage on its own. Google says its systems find the relevant piece of a long page, so you don't need to chop the content into pieces. OpenAI hasn't said how ChatGPT handles this.

Put the answer to each page's main question high on the page, in body text a crawler can read. Don't take an assistant's account of its own visit at face value, either. In the same study, two of eight assistants described their visits inaccurately.

Part 3 It picks what to use

Why ChatGPT reads a page and doesn't cite it

ChatGPT leaves most of the pages its search retrieves uncited because a local answer cites only a handful of them. We captured 116 local questions in a logged-in browser on Sept. 8, 10, and 23, 2026. The search returned a median of 12 web pages per answer, while the answer cited 2, the same split we saw when we took one local answer apart. The cited share stayed close across the three dates, with 20.7% of returned pages cited in the 60 captures, 19.4% in the 40, and 21.9% in the 16.

We read the 12 as the size of one search's result list, separate from the answer's citation choices, because each of the 16 Sept. 23 captures returned exactly 12 pages. Every cited page came from the returned set, 302 out of 302. OpenAI says ChatGPT ranks search results using multiple factors and doesn't guarantee placement. It hasn't published how an answer chooses which results to cite.

The roughly one-in-five figure holds up under other ways of counting. Counting only pages the search tool returned, 23.1% were cited. Counting the additional sites shown beneath each citation as cited puts the share at 25.9%. Across these counts, the share falls between one in five and one in four.

You can see some of the pages that didn't get cited. OpenAI says the Sources button beneath each answer lists "cited sources and other relevant links," and in our captures, the underlying data for that list included pages the search returned and the answer left uncited. Business websites turned returned pages into citations at a higher rate than directories did. Across 595 answers from our census scraper where both returned and cited pages were visible, directories accounted for 39.8% of returned pages and 29.9% of cited pages.

In our census of answers with cards, 72.5% of businesses named in the answer text appeared on that answer's card, whose rows hold listing data and aren't cited web pages, so a business can be named without any of its own pages being cited. If ChatGPT names your business while crediting a directory or another business's page, you've got a gap you can close. Open Sources on that answer, read the page that got cited, and check whether your own page gives those same facts just as plainly.

What makes a page easy for AI to quote

We think a page is easiest for AI to quote when the facts customers ask about are in plain body text, though no one, including us, has tested which page edits get a local business quoted in live AI answers. The nearest test is a 2024 lab study that found adding quotations, statistics, and source citations raised the share of an AI answer drawn from that page by roughly 30% to 40%. The study used a 2023 model, five pages chosen in advance for each question, and no local questions. It tells us about a page already selected and leaves how that page gets picked unanswered.

In our census, a business's own pages did get quoted. Among 166,017 cited links across 39,296 answers that carried a card, business websites accounted for 56.5% and directories for 29.9%. Google says its AI features don't require a special writing style because its systems understand "synonyms and general meanings," so you don't need to cover every phrasing of a question. Google also says there's no ideal page length and advises you to build pages "for your audience, not just for generative AI search."

We'd start with pages that answer what a customer wants to know just before phoning. In the body text, cover your licensing and insurance status, what drives the price, the scope of the job, and how long it takes. Include price ranges you're prepared to stand behind. Whether publishing prices changes AI answers hasn't been tested by us or by any study we've found.

How much AI favors newer content

AI assistants cite pages that are slightly newer on average than pages in Google's regular results. Still, the average cited page is almost three years old. Across about 17 million citations, Ahrefs found a 25.7% freshness gap, with ChatGPT leaning toward newer pages the most. A later Ahrefs study found that, within a single answer, ChatGPT leaned toward more established pages.

Our census scraper found that ChatGPT appended a year to 5.2% of 1,347 of its own search queries. In our logged-in captures, 28 of 111 web searches ended with an extra number (30 in 26 searches, 3650 in one, and 7 in one). The number resembles a time window measured in days, though the data doesn't label it. We recommend keeping your key pages current and dating real updates, while leaving the date alone when a page's content hasn't changed.

Does schema markup help me show up in ChatGPT?

We've seen no evidence that schema markup, code that labels facts such as your hours and address for search engines, helps a business appear in ChatGPT. Google says that for its AI features, "there's no special schema.org markup you need to add."

Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026 against 4,000 matched pages. Over the following 30 days, ChatGPT and AI Mode citations showed no real change. AI Overview citations fell 4.6% against the matched pages, a drop Ahrefs called small but statistically significant. The pages in the study were already getting cited, and Ahrefs provided no local breakdown.

Structured data still earns rich results in Google's regular search, and LocalBusiness markup can give Google details such as hours and departments. FAQ markup no longer earns a local business rich results. Since August 2023, Google has limited FAQ rich results to "well-known, authoritative government and health websites."

You don't need to remove existing FAQPage markup. Google says structured data it ignores "does not cause problems for Search."

For ChatGPT's card, we see an accurate listing as the lever because the card's rows carry listing data. Accurate LocalBusiness markup on your site is housekeeping, with hours, phone, and address that agree with the listing. We haven't found a public test of whether ChatGPT reads JSON-LD when it fetches a page. OpenAI's publisher FAQ covers crawler access and link tracking without mentioning structured data.

If you add markup, pick the type that matches your trade. Google's LocalBusiness guide advises, "Use the most specific LocalBusiness sub-type possible." It also says to list several types when a business offers several services. Schema.org's home and construction types include Plumber, Electrician, HVACBusiness, RoofingContractor, GeneralContractor, HousePainter, Locksmith, and MovingCompany, while Dentist sits directly beneath LocalBusiness.

Google requires only a business name and address, and recommends a phone number, opening hours, a location URL, map coordinates, and a price range. Your name, address, phone, and hours should be an exact match for your listing. Google recommends rating and review markup only for sites gathering reviews of other businesses, so don't mark up your own reviews.

Where links and mentions still matter

Links and mentions still count when an AI engine builds answers from search results, including for local questions, though no one has directly measured their effect on local AI answers. Google says SEO best practices still apply to its AI features. In Part 2, our Brave study found that a business's Brave rank tracked whether Claude cited it. After examining 1.4 million ChatGPT prompts, Ahrefs concluded that a page has to rank among the search results ChatGPT pulls before ChatGPT can cite it.

Mentions tracked AI Overview visibility more closely than backlinks in the one large study we found. Ahrefs studied 75,000 larger brands and found a correlation of 0.664 between web mentions of a brand and AI Overview visibility, compared with 0.218 for backlinks. Those are correlations, the study left out smaller sites, and Google warns that "seeking inauthentic "mentions" across the web isn't as helpful as it might seem." Google also says its AI features can surface commentary on a business from blogs, videos, and forum discussions.

Keep the description of your services, service area, and what you want to be known for identical on your website, your listing, and every directory that carries you. Models retain a fact better when more documents repeat it, but no AI firm claims to count agreement across sites, so consistency is good hygiene with no known ranking rule behind it. We haven't seen local data on press releases. To earn mentions, start with directories and local best-of lists already cited next to your competitors, including those covered in the directory section below.

Reddit and YouTube

For short "best plumber in [city]" questions, ChatGPT rarely shows Reddit, and we have no YouTube data. In our logged-in captures of 96 short questions, search returned a Reddit page twice, and neither answer cited it. In 20 longer first-person questions, including one about a homeowner whose air conditioner keeps running nonstop, Reddit showed up for five of the questions and was cited in two answers, though we treat that small sample as direction only. A September 2026 study of ChatGPT retrieval across many question types found that ChatGPT pulled Reddit mostly for questions seeking people's opinions and cited Reddit pages at about the same rate as other pages it read.

Our larger datasets disagree. In our census of local answers, Reddit was a source on 11.3% of 41,160 answers across seven phrasings, some long, gathered through a scraper. In our August cross-engine run, Reddit was a source on 2.1% of 2,560 ChatGPT answers. Until we reconcile those results with the logged-in captures, we don't attach one Reddit number to local answers.

The circulating claim that roughly one-third of what ChatGPT retrieves comes from Reddit traces to a 2025 study. In captured traffic, Reddit accounted for 16 of 60 spots (27%), while its citation share came from a separate dataset, so the figures describe different answers. Google and OpenAI each license Reddit's data, and Google's announcement says Google may train its models on it.

We haven't measured YouTube in local answers. Google lists videos among places its AI features can surface what people say about a business. Our view is that threads and videos help most when customers explain a problem in detail, because short "who should I hire" answers carry the card and cite directories much more often than Reddit.

Does being on Angi, HomeAdvisor, or Yelp help me show up in ChatGPT?

Angi, HomeAdvisor, and similar directories are among the sites ChatGPT cites most often in local answers, though we haven't measured whether a listing on any one raises your odds of being named. We collected 41,160 answers in the main arm of our September census from Sept. 3 to 7, 2026. Sources included Angi in 16.5% of answers, Expertise.com in 12.8%, BBB in 12.3%, Reddit in 11.3%, HomeAdvisor in 8.7%, Birdeye in 7.7%, and Houzz in 5.4%. Yelp's own site was almost never a source in those answers, though its ratings appeared on the card itself, as the star ratings section shows.

Directories also differ sharply in how often a page returned by search becomes a citation. For each directory, the table shows how many answers in our Sept. 8 captures returned its pages and how many went on to cite them.

Across all 116 captures, Thumbtack and Expertise.com pages were each returned in 45 answers, with Thumbtack cited in 2 and Expertise.com in 24. The counts show which returned pages got cited, though they don't explain why. Expertise.com runs local "best of" lists, and list-style pages continue to earn citations on local questions, even though a public study found ChatGPT citing far fewer listicles on business-software questions.

Across all five engine surfaces in our August 2026 tests, Angi, BBB, HomeAdvisor, and Expertise.com ranked among the most-cited sources (Part 2). The way someone asks also changes the results. Asking for a locally owned company while ruling out directory sites cut directory citations from 94.9% of answers to 46.1% (Part 1).

For your business, the directories and pages cited for questions about your trade in your city make a practical where-to-be-listed list. What those pages say about you gives you a second list of things to fix. Working through those lists is part of the work we do for clients.

Where ChatGPT's star ratings come from

ChatGPT's card stars match Yelp more often than Google, though Yelp never reaches a majority. We compared businesses listed on both platforms, counting a match when the card rating was within 0.05 stars of a platform's rating and the review count agreed under one of three separately applied rules: an exact count, a count within 10%, or whichever platform's count was closer. In panels of 479 businesses and 72 hand-checked businesses, Yelp matched 48% to 54% of cards and Google matched 26% to 41%, depending on the count rule. Most other cards matched neither platform, which we think mostly reflects review counts changing between card collection and our later comparison.

Each card row has a small logo field identifying the platform whose rating it displays. Of 16 rows carrying Yelp's logo, 15 had the same rating and review count as Yelp, while none of six rows without that logo did. Yelp's logo was rarer on what a customer sees first, appearing on 27.9% of 215 rows visible before Expand and 52.1% of 390 rows hidden behind it.

In the 479-business panel using the exact-count rule, cards for businesses with five or fewer reviews matched Yelp's rating 71% of the time and Google's 15%. For businesses with over 100 reviews, those shares were 22% for Yelp and 30% for Google. That pattern fits OpenAI's licensing deal with Yelp, covered in Part 1. In our Sept. 8 and Sept. 23 captures, every link to request a quote on Yelp had the partner tag utm_medium=feed_v2 in its web address.

Across 276,628 card rows in our census, the median business had 4.9 stars and 28 reviews. A Yelp profile with a solid rating often appears in ChatGPT even if you never work on it. As review counts grow, Google's figure takes over. We haven't tested whether wording in reviews appears in answers or whether higher ratings or more reviews improve a business's chance of being named.

Rating sources are visible on other engines, too. In our August cross-engine run, 97.7% of Perplexity's 265 answers quoting a rating or review count cited Yelp, while none of ChatGPT's 833 answers quoting one cited yelp.com. That ChatGPT figure covers cited web pages alone and excludes ratings displayed on its card.

Part 4 It writes the answer

How do I get my business onto the map in ChatGPT?

Keep your listing accurate, with an address inside the city customers ask about and your trade set as the primary category. We think that's how a business gets onto ChatGPT's card, though nobody outside OpenAI, us included, knows what puts a business there. The card appeared in 95.4% of 27,405 local answers in our census, with rates ranging from 92.5% to 98.2% by trade. The card doesn't depend on web search, since 95.5% of 3,303 answers still had one with the scraper's web search setting off, as we found in Part 1.

One card holds three groups of businesses. A few show before Expand, about ten appear on the full list including Expand, and some are named only in the answer text. In 57 captures with a card from Sept. 8, 2026, we counted 215 visible rows, just under four per card, along with 390 behind Expand and 46 businesses named only in text across the three lists behind one card.

We read the payload's first list as the visible card because it matched what we saw on the rendered page for the one question we checked by hand. Five businesses appeared before Expand in that check. DataForSEO's scraper, which we used for our census and which is widely used across the industry for measurement, combines the visible card and Expand pool into one list. Tools built on it report that combined pool, while the five businesses visible first in our hand check aren't separated out.

The business details on the card track Google's listing data, though the data doesn't identify its provider. Keep your Google listing right. Gemini behaved differently in our August run. Gemini showed no card or other places layer in its app or API, even though its retrieval runs through Google.

Each card row has fields you don't see, including whether the listing is claimed and a field called ranking_score. That score appears on every business record we captured, but the card's display order doesn't follow it. The field alongside it, popularity_score, has been blank on every record. Neither field is documented, and we can't say what gets a business onto the card until we've modeled them.

In OpenAI's documentation, the routes we found for feeding a local business's details to ChatGPT are partner programs covering quote requests and restaurant reservations. OpenAI's Get Quote spec says the plugins are "currently in beta and being tested with approved partners," who apply through the ChatGPT merchants form and submit a feed of the businesses they list. It also says, "ChatGPT indexes these records for search."

In 16 captures dated Sept. 23, 2026, every card record carried the feed's quote field, service_providers, among its 57 fields. That field was filled for 73 businesses in those captures, and every one named Yelp as the provider with the quote widget described in the spec. We expect most owners to reach this route through a platform with an existing plugin, and in all our captures that platform is Yelp.

The card isn't the only way to be named in an answer. Among answers with a card, 72.5% of businesses named in the text were on that answer's card, leaving about a quarter named without a card spot.

How does ChatGPT decide which businesses to recommend?

ChatGPT names businesses from a local pool that shifts from one ask to the next. OpenAI hasn't published how it selects names from that pool. In our census, 99.7% of 27,405 local answers named at least one business, and each answer named an average of 6.6 businesses.

In an August 2026 repeat study, we asked 40 local questions 240 times each over three days. For the average question, about 28 businesses appeared in at least 5% of its answers. The median business in a pool appeared in 30% of answers, while just 36 of 1,115 business-and-question pairs appeared in 95% or more. For a typical business in the pool, that works out to getting named about three times in ten.

The pools are local. Across 44,988 answers in our census of who gets named, ChatGPT named 88,186 distinct businesses, and 85% of a coded sample of 238 operated under their name in only one metro, with a 95% confidence range of 81% to 90%. Only 11.7% of businesses shown on ChatGPT's local cards appeared on cards in more than one of our 24 metros, which we read as good news for small independents whose competition is mostly other local businesses. In a random sample of 600 answers to a request for locally owned businesses that ruled out chains, every answer named at least one business we could confirm exists and that fit the request. Wording changed the average count from 3.4 names for "list three" to 10.6 for the locally owned request (Part 1).

We haven't tested whether ChatGPT uses Google's public local-ranking factors. Google's local-ranking help page says, "Local results are mainly based on relevance, distance, and popularity," and under its heading for the third factor, prominence, it counts "how many websites link to your business and how many reviews you have." In our Sept. 8 and Sept. 23 captures, distance_meters was blank on every card record, and the adjacent ranking_score didn't match card order. Our data neither shows nor rules out ChatGPT weighing distance or reviews as Google says it does. The Part 2 census rows fit relevance and distance, with most card businesses having an address in the principal city and listing the asked trade, though those rows don't establish why a business made the card.

In 70.4% of answers with a card, at least one named business was absent from it. If competitors get named and you don't, you can check whether you're absent from the card, whether the pages the answer cites leave you out, whether AI crawlers can't read your site, or whether the answer varied between asks. Part 6 puts those checks in order.

Does ChatGPT show ads for local businesses?

Yes, ChatGPT shows ads on its Free and Go plans. On Sept. 23, 2026, we saw ads on 10 of 16 fresh local questions from our free account. OpenAI began testing ads in the US on Feb. 9, 2026. OpenAI says the ads sit below the end of a response with a sponsored label and don't appear for Plus, Pro, Business, Enterprise, or Edu accounts or in Temporary Chats.

OpenAI says ads "do not influence ChatGPT's answers." It says ads can't appear next to sensitive or regulated subjects such as personal health and mental health, and health advertisers face stricter eligibility rules. You can sign up at ads.openai.com if it's available in your country, and OpenAI's advertiser guide suggests a starting bid of $3 to $5 per click. If you advertise, let OAI-AdsBot through because OpenAI says the bot visits an ad's landing page to review it for policy compliance and judge when the ad should show.

Lead marketplaces bought most of the 10 ads we caught on Sept. 23. Three came from Thumbtack and three from Angi, while local businesses and brands accounted for the other four. One of the 10 advertisers also appeared on the card in the same answer. Thumbtack appeared in ChatGPT's search in 45 of 116 captured answers and was cited in two.

Ad tracking tools may miss these ads entirely. Hours later, a commercial scraper found no ads for the same 16 questions. On Sept. 8, the same account saw no ads across 63 captured answers. Our Sept. 23 count comes from one free account on one evening, so it can't tell you how common ads are on local questions overall or on other plans.

How to fix wrong information ChatGPT gives about your business

Correct your Google listing and Yelp profile first, then your own site. The business details on a ChatGPT card follow Google's listing data, while its star ratings come from Yelp more often than from Google.

We found no direct way for a business to correct its own details through OpenAI's help center. You can still flag a wrong answer with the thumbs-down under it and choose from the issues it lists. OpenAI's reporting page says its Model Quality team may review reported domains and content and apply filters "to help prevent ChatGPT from relying on unreliable sources in future responses." That page addresses content that may breach OpenAI's terms or the law, and neither route is presented as a way to correct business details, so we'd use them alongside the listing fixes.

Tap your business on the card and check the details that appear, including its phone number, address, hours, and open-now status. The row for your business also records whether the business is permanently or temporarily closed and whether its listing is claimed.

Check every verifiable claim ChatGPT makes about you beyond the card, including claims about your locations, staff, and services. In our checks, we saw ChatGPT extend a business's service to an adjacent service the business didn't offer. We also saw it invent a staff member, and both mistakes showed up often enough to watch for. We haven't measured how often either mistake happens.

Some businesses ChatGPT names couldn't be traced through the sources we checked. In a random sample of 600 answers to a locally owned question, 17.3% named at least one business we couldn't trace through ChatGPT's map cards, the business's own website, a Google listings database, or Google Places. That figure is an upper bound because a sole proprietor with no listing could fall into that group too.

Pay attention to the descriptive words ChatGPT attaches to your business, such as "affordable" or "family-friendly," and check them again every few months. Those words are part of how it presents you to customers, and our goal is to narrow the distance between that description and the one you'd want. Keeping your hours, phone number, and services consistent across your site, listings, and directories is good hygiene, though no AI firm claims to count that cross-site agreement. We also haven't measured the error rate for phone numbers and hours on cards.

Part 5 You measure it

Why does ChatGPT give a different answer every time?

Every ask is a fresh sample from a broad pool, so the same question seldom brings back an identical list twice. In our repeat study, two runs of the same local question a few minutes apart named the identical set of businesses 3.6% of the time across 188,520 pairs of runs. We collected those runs through a commercial scraper of the ChatGPT app in August 2026.

A language model builds answers piece by piece, and Thinking Machines Lab got 80 different answers from 1,000 identical requests to an open model even though it was set to always pick its likeliest next word. The lab tied most of that spread to the number of other requests hitting the server at the time. The web search underneath shifts too, with sites that two runs in one sitting had in common averaging 46% of all sites either returned when both kept their search results. The broad pool in our study held roughly 28 businesses per question, and the median business appeared in 30% of answers, as the section on how ChatGPT picks businesses discusses.

More asks in one sitting won't settle the list. Across 6,624 blocks of five back-to-back runs for businesses ChatGPT named about half the time, a business appeared one to four times in 89% of blocks, missed all five in 6%, and appeared all five times in 5%. Within a sitting, each run acted like an independent draw at that sitting's own rate, while between sittings 12 or 24 hours apart, the rate itself shifted. For a business with a 5% to 20% naming rate, runs beyond about 65 in one sitting added nothing.

Two scans of identical, untouched questions had to differ by about 28 points before the change cleared the answers' movement between sittings, however many runs went into each scan. That 28-point threshold for calling a change real comes from 72 midweek hours, so movement over a month can only be larger. Spreading asks across days lowers the bar. For a business with a 20% to 50% naming rate, two two-week windows of daily asks with about 70 answers apiece need a difference of about 15 to 18 points, based on the study's estimate of movement between sittings.

How someone asks changes how many businesses an answer names, too. We track several phrasings per question over many days and report how often a business gets named, with a margin of error. We never report a position. The full method is in How to run AI visibility tracking.

How to check whether ChatGPT names your business

Ask ChatGPT questions your customers would ask in a few different wordings over several days, then tally how often it names your business. Searching your business name tells you what ChatGPT knows about you. It doesn't tell you whether ChatGPT recommends you.

Test three kinds of question separately, since you may do well on one and poorly on another. A broad trade-plus-town question checks whether you get recommended at all. A question about a specific service checks whether you get recommended for the jobs you want. A question naming your business shows how ChatGPT describes you, and the section on wrong information can help you check that description.

Keep your name out of the first two kinds of question, since naming a business in the question gives it an advantage in the answer. Put your town in the wording instead. OpenAI says ChatGPT may infer location from your IP address, and for local results that show the wrong area, its help page says, "Include your city, neighborhood, or postal code in your question." Vary the wording of each question, too, since the wording alone changed how many businesses an answer named.

Record the other businesses ChatGPT names and the sites it cites in answers that leave you out, since those businesses hold the spot you want and those cited pages are ones you'd want to appear on. The directory section helps turn them into a list. Tracking tools run this check at scale, and our tracking guide covers assembling the question set. The route a tool uses to reach ChatGPT affects what it can see.

How accurate are AI visibility tools?

A tool's accuracy depends on its route into the engine, and for ChatGPT, tools commonly pull answers with a scraper that finds most card businesses but misses much of the surrounding detail. On Sept. 8, 2026, we asked 60 local questions through a ChatGPT-app scraper and a logged-in browser. The scraper found 208 of the 215 card businesses the browser showed, while ChatGPT's own search queries appeared in 9 of the scraper's 180 responses and in all 60 browser answers. Our tracking guide compares the routes and lists what to ask any tracker, ours included, before trusting its figures.

Google and Microsoft now report on their own AI answers, and Google's free Search Console shows how often your pages appeared in AI Overviews and AI Mode, as covered in the section on AI Overviews. In February 2026, Microsoft added a public-preview AI Performance report to Bing Webmaster Tools that measures how often Copilot and Bing's AI-generated summaries cite your pages, though neither report says it includes ChatGPT and Microsoft doesn't list which partner services its report covers. We've found no OpenAI report showing a business its own mentions or citations, as discussed in the traffic section. For ChatGPT, the free route is the manual check.

What to track to measure AI visibility

Track how often you're named by question and by engine, as a share of every answer in a window, alongside a band showing how far that share drifts on its own. We call that share visibility. It equals the number of answers that named you divided by the total number of answers in the window. Don't average daily percentages, and count a move against the prior window only when it clears the band.

Group your questions into a cluster for each service-and-town combination that brings real jobs, with several phrasings in each cluster. A single phrasing measures only one route into an answer. Give each engine its own line in the report. The engines find businesses differently, and a blended figure hides which engine you're absent from.

For a local business, track three more numbers. Record how often you appear on ChatGPT's card separately from how often you're named. Compare your listing's star rating and review count with the median for businesses shown on ChatGPT's cards, 4.9 stars on 28 reviews. Track how often your own site gets cited, since being named and being cited are distinct outcomes.

Share of voice is your share of all business names across the answers, so keep it separate from visibility and never present one as the other in a report. Rank and average position don't describe anything stable and shouldn't be targets, since the same set of businesses appears in only 3.6% of cases when the same question is run twice. A one-day change reflects one sitting, so don't use it as a target. A single composite score shouldn't be a target because it hides the gaps by question and by engine that tell you what to fix.

When you make a change, mark its ship date, compare the window before it with the window after it, and count any move inside the band as no change. The section on running this over time covers ongoing checks. Any agency that reports on mentions ought to produce each of these numbers for you, and the section on buying AI SEO lists the questions to ask. The tracking guide goes into what a report should show in more depth.

How to measure traffic from ChatGPT

Count AI-tool visits in Google Analytics 4 (GA4) by referring site, and treat the total as a minimum because many AI visits arrive unlabeled. GA4 now has an AI Assistant channel, which Google says covers visits "from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok" and "excludes Google's AI Overviews and AI Mode." In a year of GA4 data from our properties and client properties, we found chatgpt.com visits across five mediums on one site, with each channel holding only part of those visits. Our post on seeing AI traffic on your website covers building a channel to gather AI visits.

ChatGPT tags its outgoing links. OpenAI says ChatGPT "automatically includes the UTM parameter utm_source=chatgpt.com in referral URLs." In a logged-in session on Sept. 8, 2026, we checked one shopping answer. All eight brand links carried the tag, while none had a medium or campaign.

Some visits arrive bare. When an app sends no referrer and its link has no tag, GA4 places the visit in Direct, which Google defines as visits "via a saved link or by entering your URL." Clicks from AI Overviews and AI Mode register as ordinary Google search clicks, while Search Console's AI report counts their impressions. For a local business, a lot of the effect may end in a phone call. Someone who learns about you in ChatGPT and calls the number on your listing leaves nothing connecting that call to the answer.

As far as we've found, OpenAI offers businesses no report on their mentions, citations, or referrals. Its publisher FAQ covers crawler access and the tag. The only conversion measurement in its developer docs concerns ads purchased in ChatGPT.

A study of how well AI visitors convert can count only the visits it can identify. In its study of 97 business-to-business sites, Orbit Media identified AI visits by source and GA4's AI Assistant medium and noted that AI-app visits appear as Direct. We don't apply those conversion rates to local businesses because the sites are business-to-business. Bots appear separately in your CDN or server logs, and you should count only those your CDN verifies. Senders choose their own bot names, and Cloudflare's documentation marks a filter on that name "Can be spoofed."

Part 6 What to do about it

Finding your gap

Before you fix anything, confirm there's a gap, then identify its type, since each type calls for its own fix. A single ask can't tell you whether your business is absent or you caught a bad-luck sitting. Start with a baseline of your naming rate over one to two weeks of asks, using the section on checking.

Part 4 gave four reasons ChatGPT could name your competitors and leave you out, and we check them in order alongside other questions that separate one gap from another. Each question points to a step in the plan below. They're sorting questions. No published evidence shows that any one of them identifies a cause.

1. Can AI crawlers read your site? A block in robots.txt or your host's settings can shut your pages out of the web searches these answers rely on. It takes a few minutes to check (step 1).
2. Are you on the card for your trade and town? If you're absent, check your listing for an address in the town customers ask about and your trade as its primary category. The card section treats these as the first entry conditions to verify (step 2).
3. Do the pages ChatGPT cites for your trade and town mention you? Check the pages in ChatGPT's Sources list for your trade and town. That list also gives you places to get listed (step 3).
4. Does ChatGPT know your business? Ask about your business by name. If it describes you well but never names you for your trade, it knows you and passes you over, so start at steps 3 and 4. If it can't describe you, start at steps 1 and 2.
5. Which service, town, or engine is the hole in? One early client was named in 0 of 12 answers for one town and 6 of 12 for another, a split its overall total masked. An unclaimed service is easier to capture than one held by a competitor (step 4).
6. Does a competitor get credit for something you do? Ask ChatGPT "who around here [does the thing you're known for]?" repeatedly, then tally who gets the credit (step 4).
7. Are you named while another site gets the citation? Your standing then depends on pages outside your control (steps 3 and 4).
8. Did the answer just vary? If the other questions don't surface a gap, compare your windows against the band before changing anything (step 6).

How do I get my business recommended by ChatGPT?

Start with a crawler-readable site and an accurate listing, work to get onto the pages local AI answers cite, publish pages that answer customers' questions, and track how often you're named over time. We haven't found a published controlled trial of this plan on real local AI answers, and we haven't run one. Each step rests on what we've measured in local answers, so none is a tested result. Our AI search work for local businesses begins with these same checks, starting with crawler access.

1. Make your site readable Check robots.txt and your host's bot settings, allow OAI-SearchBot in, and make sure your homepage's raw HTML states your services and where you work. Roughly one in four local business sites in our access census failed that check. See Can AI crawlers read my site?
2. Keep your listings accurate Use an address in the town customers name and your trade as the primary category, and keep your phone number and opening hours current on Google and Yelp. Review your Yelp profile, since the stars on the card come from Yelp more often than Google. Microsoft says a Bing Places registration helps your address, hours, and contact info stay current and "eligible for inclusion in AI-generated responses." See the card section and the star ratings section.
3. Get onto the pages these answers cite Use the Sources lists on trade-and-town answers as your target list. Get listed on pages where you're absent, and correct any inaccurate details those pages carry about you. See the directory section.
4. Publish pages that answer what customers ask before they call Cover every service and town you serve, what drives prices, licensing and insurance, and the specialty you're known for. Put the answers in plain body text, including price ranges you'll stand behind. See What makes a page easy for AI to quote?
5. Keep markup accurate Use the LocalBusiness type that fits your trade, with a name, address, phone number, and hours identical to your listing. Treat it as housekeeping. See the schema section.
6. Measure and log changes Track how often you're named across phrasings and days, with a band for how far the rate moves on its own, and log the ship date of each change. See What should I track?

The order stays the same whether you're a plumber, dentist, law firm, or restaurant. What changes by trade are the directories that get cited, the right schema type and, for some professions, the rules about requesting client reviews.

Is AI SEO worth paying for?

Of course it's worth paying for. People asking ChatGPT for someone to call are done comparing, and the businesses it names receive the call. Local SEO is the foundation, and AI search adds a layer with its own sources and its own scorekeeping. The engines differ in what they cite, as the section on how each engine finds businesses shows. ChatGPT cited Yelp on none of 2,560 answers, while Claude cited it in a third of its answers. Our access census found roughly one in four local business sites can't be read by at least one AI crawler. Plenty of local businesses haven't started on the AI search layer, so those that begin now can get named while competitors are still catching up.

What you pay depends on how much an agency does on your behalf, so compare quotes by what each includes. A useful audit gives data-backed answers to the gap questions above, including a baseline rate for each engine with a band. The questions we'd ask of any tracker work for an agency too:

• Does each engine's figure come from the app itself, a scraper, or an API?
• How many distinct days and answers sit behind each figure?
• Do you show a band and report movement inside it as no change?
• What counts as being named, and what's the miss rate of that check?
• Does "on the card" mean the handful a customer can see or the full pool behind Expand?
• Do you report a rank or an average position?
• Which of my listings, directory profiles, and pages will you change, and how will you record the date each change goes live?

Choose an agency by its method, and don't use ChatGPT's own agency recommendations as a vetting signal. We asked it "best AI SEO agency for local businesses" in three separate sittings and got 13 namings covering 11 agencies, with none named in all three. For "best GEO agency for small business," six of 12 namings across three answers rested on a "best agencies" list published by the agency named, and each answer warned about agencies ranking themselves while still naming that list's publisher. Check reviews hosted outside the agency's control and ask to see a client's figures with a band attached. Our AI SEO work logs every change to the month it went live without reporting a position inside an answer.

Who runs it, and how often to check

One person, whether that's you, a staff member, or an agency, should own the question list, the listings, and a change log, then read the numbers monthly over a rolling window. Our tracker asks each phrasing once a day per engine and reads a 30-day window. The client report, which we're still building, compares each calendar month with the prior month and includes the band.

Keep question wording fixed once tracking starts, since new wording starts a new series, and log the ship date for each change so the chart shows before and after. Set goals as rates with a band. A rank, a single day's change, and one blended score make poor goals for the reasons in What should I track? Recheck the listing and cited pages after you move, add a service, or change your hours. The full cadence is in the tracking guide.

Getting ready for AI agents

For most local businesses, preparing for AI agents isn't urgent. There's still one small step worth taking now. Google describes AI agents as systems that "perform tasks on behalf of people, such as booking a reservation," and points readers to its agent-friendly site guide when agents matter to your business and "you have extra time."

The local-business actions we've seen in ChatGPT run through partners, and OpenAI's Get Quote plugins are in beta and limited to approved partners, as the card section explains. Restaurant table bookings in ChatGPT go through OpenTable, Resy, and Yelp. Agent standards are still early, and WebMCP, which lets a site offer agents its own tools, "is not a W3C Standard." Signed requests let a site confirm an agent and depend on IETF drafts, while agent commerce protocols so far are mostly designed around product purchases.

For now, make your quote and booking forms easy for agents to use. OpenAI says its Atlas browser agent relies on ARIA labels and roles, which screen readers also use. Clear labels on buttons and form fields serve both.

Methods

Numbers this page publishes for the first time rest on our own data. Each note says where one set of them comes from.

• Concept questions. 92 questions about AI search, written the way an owner would type them, asked in fresh chats on one ChatGPT Plus account in three separate sittings on Sept. 24 and 25, 2026, with the searches read from the conversation data. Counts from the first sitting are one run per question, and comparisons across sittings use all three. Audited Sept. 24 and 25, 2026.
• Logged-in captures. Conversations captured from a logged-in free ChatGPT account with memory off on Sept. 8, 10, and 23, 2026, 116 in all, with the search call readable on 100. Each lookup was compared word for word against the typed question, and every distinct page the search returned and the answer cited was counted. The number at the end of some web searches carries no label in the data. Business records come from the 16 captures of Sept. 23, counting each business once per capture. Audited Sept. 24 and 28, 2026.
• Reddit counts. From the same 116 captures, split into 96 short "best or who does X in [city]" questions and 20 long first-person ones, and from the August cross-engine run. Small samples, read as direction.
• Census. The pre-registered analysis of 44,988 answers collected Sept. 3 to 7, 2026 through DataForSEO's ChatGPT scraper. The web-search-off arm is 3,303 answers to the plain question with the scraper's web search setting off, and each phrasing comparison uses 3,915 answers on matched trade and metro pairs. Named-business counts use our extractor, with 92.6% recall and 94.1% precision on 150 answers coded blind by two AI coders and read a third time.
• Repeat study. 12,120 ChatGPT answers to 46 questions, 40 of them local, collected through a commercial scraper in seven bursts 12 hours apart from Aug. 24 to 27, 2026. Pre-registered. A business enters a question's pool when it appears in at least 5% of that question's answers. Every across-day figure is a floor on month-to-month movement. The 15 to 18 point figure is our arithmetic, using the study's figure for movement between sittings.
• Cross-engine run. The same 10 local prompts sent to ChatGPT through DataForSEO's scraper, to Claude and Perplexity through their APIs, and to Gemini through its API and its consumer app, from Aug. 24 to 27, 2026. Shares count answers that cite each site. The scraper doesn't record ChatGPT's licensed Yelp data as a cited source.
• Rating matching. Card businesses listed on both Yelp and Google were matched on star rating, within 0.05 stars, and on review count under three rules (exact, within 10%, and nearest), on a coded panel of 479 and a hand panel of 72.
• Existence check. Every business named in a random sample of 600 answers to the locally owned phrasing was looked up in ChatGPT's map cards, on its own website, in a database of US Google listings, and in Google Places. Untraceable is an upper bound.
• Access census. 24,000 independent local business websites sampled from Overture Maps listings in 40 US metros and tested on Aug. 20, 2026 with user agents that say they're a test. Unreadable means blocked by robots.txt, challenged at the host or CDN, or missing basic business details in the raw HTML. Host and CDN figures are lower bounds.
• Scraper comparison. The same 60 local questions asked on Sept. 8, 2026 in a logged-in free ChatGPT account with memory off and through a commercial scraper, three scraper runs per question. Audited Sept. 8 to 10, 2026.
• Analytics and client figures. The GA4 medium count comes from our own analytics data, checked Sept. 28, 2026, and the tagged-link count from one browser session on Sept. 8, 2026. Client figures were rebuilt from our tracking database on Sept. 25, 2026.
• Agency questions. Two hiring questions asked in three separate sittings on one ChatGPT Plus account on Sept. 24 and 25, 2026, with each agency traced to the pages cited beside it in the answer.
• OpenAI's documentation. Checked against its help center and its developer documentation index on Sept. 25 and 28, 2026.