How to check if AI assistants recommend your business

Open a fresh, logged-out AI chat. Ask what a customer would ask — “best [your service] in [your city]” — and run it three times in three separate chats. Write down whether you were named, who else was named, and which websites the AI cited. Then check your robots.txt to confirm AI crawlers can actually reach your site.

That is the whole method. It takes about 30 minutes and costs nothing. The rest of this guide is how to do each step properly, and — more importantly — how to read the results without fooling yourself.

Why this is suddenly worth 30 minutes

Consumers moved first. BrightLocal’s 2026 Local Consumer Review Survey found that 45% of consumers had used an AI tool to find a local business in the past year, up from 6% the year before. That makes AI the third most-used local discovery tool, behind Google and Facebook, and ahead of Yelp and TripAdvisor.

BrightLocal, Local Consumer Review Survey 2026 — 1,002 US adult consumers, published March 10, 2026. The same survey found the use of Google reviews slipping from 83% to 71% year over year. That is review-reading behaviour specifically, not Google’s overall share of local discovery — a distinction worth keeping, because the looser version of this stat gets quoted a lot.

AI recommendation is far more selective than Google. SOCi’s 2026 Local Visibility Index analysed roughly 350,000 business locations across 2,751 brands. Those brands appeared in Google’s local 3-pack 35.9% of the time on average. ChatGPT recommended only 1.2% of the locations. Perplexity managed 7.4%, Gemini 11.0%.

SOCi, 2026 Local Visibility Index, via Search Engine Land. One caveat most articles quoting the 1.2% leave out: SOCi studied multi-location brands — franchises and chains. A single-office practice is not in that dataset. The number tells you AI recommendation is dramatically more selective than Google, not that your own odds are 1.2%. Which is exactly why running the check yourself beats reading a statistic.

The practical takeaway is the same either way. Where Google shows ten links and lets the customer choose, an AI assistant names two or three businesses and the customer usually stops there. Being on page one is no longer the same thing as being in the answer.

The mistake that invalidates most self-checks

Almost everyone who tries this does it wrong in the same way: they ask once, screenshot the result, and draw a conclusion.

That does not work, because AI answers are not stable. SparkToro and Gumshoe.ai had 600 volunteers run 12 brand-recommendation prompts a combined 2,961 times through ChatGPT, Claude, and Google’s AI search. They found less than a 1-in-100 chance that ChatGPT or Google’s AI would return the same list of brands on any two runs of the same prompt. For the same list in the same order, it was closer to 1 in 1,000.

SparkToro / Gumshoe.ai. Claude scored marginally better than ChatGPT and Google on list consistency, but still poorly.

Three setup rules that matter more than they sound:

• Use a logged-out or temporary chat. If you are signed in, the assistant may remember previous conversations — including the ones where you talked about your own business. You will get a flattering answer no stranger would ever see. Use ChatGPT’s Temporary Chat, or just a private browser window.
• Start a genuinely new chat for each run. Asking the same question three times in one conversation does not give you three independent samples. It gives you one answer and two follow-ups shaped by it.
• Name your city explicitly. These tools guess location from your IP address, so testing from your office can produce results a customer across town would never see. Write “in Scottsdale, Arizona,” not “near me.”

Step 1: Write the questions your customers actually ask

Not the keywords you would bid on. The sentences a real person types when they have a problem and no shortlist.

Six queries × three runs = 18 chats. That is your 30 minutes.

One thing worth knowing as you build the list: the type of question changes how stable the answer is. Conductor ran 14,000 API calls across 10 industries, seven intent types, four models, and five personas, and found comparison prompts the most consistent intent type by a wide margin — 63% brand overlap between runs, with the same brand leading 91% of the time. Recommendation prompts managed 49% overlap; purchase-intent prompts came last at 40%.

Conductor — the ranking held across all four models and all ten industries. Expect your problem-first and superlative queries to bounce around more than your comparative ones. That is the medium, not a measurement error.

Step 2: Run them on more than one assistant

Run your list on at least ChatGPT and one other assistant. Perplexity and Google’s AI Mode are the useful comparisons.

Do not assume results transfer between them. There is no shared local index. Each assistant grounds its answers in different sources, and a business that dominates one can be absent from another. This is one of the most common and expensive misunderstandings in the category.

It is also an area where the conventional wisdom has gone stale. For most of 2025 the standard advice was “ChatGPT runs on Bing, so fix Bing Places.” That was reasonable at the time — ChatGPT Search was built on a Bing-derived web index. But independent testing through 2026 has found ChatGPT’s local grounding leaning far more heavily on Google-sourced data than that advice implies.

One May 2026 analysis of 1,732 citations across nine cities found google.com outweighing bing.com by roughly 130 to 2 in local answers. Treat this as directional, not settled: which listings feed the answer is now an empirical question, and the sources panel in your own results is better evidence than anything you will read in a blog post — including this one.

Step 3: Score what you got

Keep it in a spreadsheet, one row per query, four columns:

1. Named? How many of the three runs named you. Record it as a fraction (2/3), not a yes/no.
2. Position. When named, were you first, or fifth in a list of six? First matters enormously; buried in a list of ten barely registers.
3. Who else appeared. Every competitor named, every time. This is the most useful column in the sheet.
4. Sources cited. The diagnostic one — more below.

Then compute one number: across all 18 runs, in how many were you named? That is your mention rate. It is a rough estimate with a wide margin of error at this sample size, but it is a real baseline, and re-running the same list in 90 days tells you whether anything moved.

Reading the competitor column

If the same two or three businesses appear in nearly every run, you are in a consolidated market. Those competitors have accumulated enough consistent signal across the web that the model is confident about them, and dislodging them takes months of foundation work rather than a quick fix.

If the names change constantly between runs, the category is unsettled. Nobody has locked in the position, and you are competing against noise rather than an incumbent. That is the better situation to be in, and the one where consistent work pays off fastest.

Step 4: Read the sources — this is where the answer actually is

Most people stop at “was I mentioned.” The sources panel is worth more than the answer text.

Expand the citations on each response and write down every domain. After 18 runs you will have a list of the sites actually feeding recommendations in your category and city. Typically some combination of:

• Business websites — often the single largest category
• Vertical directories — Avvo or Justia for legal, Healthgrades or Zocdoc for medical, Angi for home services
• General directories and review platforms
• “Best of” and comparison listicles, often from small local publishers
• Local news and city guide sites
• Sometimes Reddit or YouTube

This matters because most of what feeds an AI recommendation is not on your own website. Yext’s analysis of 6.8 million AI citations found 86% came from sources a brand already controls — its own site (44%) plus its listings (42%). That is the good news. The bad news is that “your listings” means a specific, discoverable set of platforms, and almost nobody knows which ones they are until they look.

Yext — 6.8M citations from 1.6M queries per model across ChatGPT, Gemini, and Perplexity, over July and August 2025. The same analysis put Reddit and similar forums at just 2% of citations once location and query intent were applied.

Step 5: Check that AI crawlers can reach your site

This takes two minutes and is the one step that occasionally uncovers something catastrophic. Type your domain followed by /robots.txt into a browser. Look for any block on these user agents:

• GPTBot and OAI-SearchBot (OpenAI)
• PerplexityBot
• ClaudeBot
• Google-Extended
• Applebot-Extended

A line reading Disallow: / under any of those names means you have told that system to stay out. Sometimes a developer added it deliberately years ago. Sometimes a plugin or a template did it and nobody noticed.

This is worth checking rather than assuming, because the defaults have changed. On July 1, 2026 Cloudflare replaced its single “block AI bots” switch with three categories — Search, Agent, and Training. As of September 15, 2026, Training and Agent crawlers are blocked by default on pages that display ads, while Search crawlers stay allowed. Those defaults apply to new customers, new sites added by existing customers, and all existing free-plan customers.

Cloudflare, “Your site, your rules: new AI traffic options for all customers”.

Two caveats so you do not panic unnecessarily: the default only fires on ad-displaying pages, which most local service sites do not have, and existing zones with an explicit saved preference are respected. But there is a real trap in the fine print — Cloudflare evaluates multi-purpose crawlers under all of their behaviours, so blocking “Training” can also catch Googlebot, Bingbot, and Applebot. A blanket block aimed at AI can quietly damage your ordinary search rankings.

If you are on Cloudflare, open Security → Settings and find the AI bot policies. Check which one your zone is actually on rather than assuming — an inherited default is easy to miss, and it is the setting most likely to be blocking you without anyone having decided to.

What your results mean

Named in 0 of 18 runs. The common outcome, and less alarming than it feels. Work through the source list from Step 4 in order of how often each domain appeared. Claim what you can claim, correct your business name, address, and phone everywhere they are wrong, and re-measure in 90 days. Expect the first movement to show up as your listings getting cited before your name appears in the answer text.

Named in 1–5 of 18. You are in the consideration set but not confidently. The model knows you exist and is not sure enough to name you every time. Consistency of information across platforms is usually what is missing — conflicting hours, an old suite number, a name with “LLC” in some places and not others.

Named in 6+ of 18. You are doing well, and the job changes from getting in to staying in. Re-run this list quarterly. Watch the competitor column: someone else’s push shows up there before it shows up in your own mention rate.

Three things not to spend money on afterward

The GEO space has filled up with paid services fast, and some of what is being sold has been tested and found wanting.

llms.txt files

Ahrefs analysed 137,210 domains and found 97% of published llms.txt files received zero requests in May 2026 — no bots, no humans. Of the traffic the rest did get, 96% came from bots, and the largest identified slice was SEO audit tools checking whether the file existed. AI crawlers did not even probe for the file on domains that lacked it. It costs half an hour to publish and there is no known penalty, so ship one if you like — but do not pay anyone for it, and do not expect citations from it.

Ahrefs — domains in Ahrefs Web Analytics with traffic in May 2026.

Schema markup sold as an AI-citation strategy

Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026 against 4,000 matched control pages. The result: no meaningful citation uplift on Google AI Mode (+2.4%) or ChatGPT (+2.2%), both small enough to be indistinguishable from noise, and a 4.6% decline in AI Overviews that was measurable against controls — though Ahrefs itself stops short of pinning that decline on schema.

Ahrefs. Read the scope carefully: every page in the study was already heavily cited before schema was added — the dataset required 100+ AI Overview citations beforehand. It shows that bolting schema onto an already-visible page adds nothing. It does not show that schema is useless for an unknown local business trying to establish what it is and where. Keep valid LocalBusiness markup for ordinary search reasons; just do not buy it as an AI visibility product.

“AI directory submission” packages

Bundles of dozens or hundreds of directory listings, sold on volume. Your Step 4 source list already tells you which handful of platforms actually get cited in your city and category. Fifteen right ones beat two hundred random ones, and the difference is measurable in your own spreadsheet.

Do it again in 90 days

The most useful thing about this exercise is not the first result. It is having a baseline to compare against.

Save the spreadsheet. Save the query list exactly as written — changing the wording later invalidates the comparison. Run it again in 90 days and look at three things: whether your mention count moved, whether your listings started showing up in the sources panel even when your name did not appear in the answer, and whether the competitor set shifted.

That middle one is the leading indicator worth watching. Foundation work shows up in the sources before it shows up in the answer, and if you only track whether you got named, you will conclude nothing is working during exactly the period when things are starting to.

Method notes

Everything above describes manual sampling, and it has real limits worth stating plainly. Eighteen runs is a small sample with a wide confidence interval — enough to distinguish “invisible” from “sometimes present,” not enough to detect a change of a few percentage points. Results vary by location, by date, and by model version. Anyone claiming a precise AI visibility percentage from a handful of queries is reporting noise with decimal places.

The method is still worth running. A rough number you measured yourself and can repeat beats a precise number someone sold you.