The websites AI assistants actually read before recommending a local business

When a customer asks ChatGPT for a good plumber in their city, ChatGPT does not know the answer. It goes and reads a handful of web pages, then writes a recommendation based on what it found. The whole question of whether your business gets named comes down to which pages it picked up, and whether your name was on them.

So we logged them: every source cited across five AI assistants, for real local businesses in four cities. The five do not read the same websites, and the gap between them is wide enough that fixing the wrong thing first is easy to do.

How we measured this

We ran audits on five local businesses in Austin, Ann Arbor, Toronto and Mumbai, asking each AI assistant the questions a real customer would ask. That produced 115 answers, and we stored every source each one cited.

Three things about that sample. All five businesses are restaurants and cafes, so the ordering below is food-weighted and we say so wherever it matters. Two of the five are outside the United States, while most of the platforms discussed here are US infrastructure, so the ordering is US-weighted too. And we count a business as named whether or not the assistant linked to it, because in local answers most mentions carry no link. When we checked Franklin Barbecue’s results, 15 of its 18 mentions included no link to Franklin’s own site. The answer named the business and moved on.

Anyone measuring this by looking for links is going to report famous local businesses as invisible.

Which websites get read the most?

Across all five assistants, these came up most often:

One answer often cites several sources, which is why these add up to more than 115.

Reddit at the top surprised us, and we think it is a restaurant effect rather than a general rule. Foundation and AirOps ran a much larger study across more than 28 million small business queries in the fourth quarter of 2025 and found Yelp far in front at 512,680 citations, roughly 3.4 times the next platform, and taking 72.5% of the citations Google’s AI Mode gave to any of the six platforms they tracked. Their next four were the Better Business Bureau at 149,700, Angi at 145,600, Thumbtack at 56,000 and HomeAdvisor at 33,600.

Our working explanation is that Reddit runs hot for restaurant recommendations, where people ask other people, while Yelp is stronger across services where people look for ratings and licensing. That is a hypothesis, not a finding. We are running the same checks on trades now, and it will either hold or it will not.

Does every AI assistant read the same websites?

No, and the split is wider than most owners expect.

A business with a thin website and a strong Yelp page will get named by Claude and skipped by ChatGPT. A business with a detailed website and no reviews anywhere will get the opposite result. Neither owner is doing anything wrong, and nothing on their end tells them which half they are missing.

The rule that falls out of this is worth keeping: your own website is what reaches ChatGPT, your directory listings are what reach Claude, and your reviews and mentions elsewhere are what reach Gemini and Perplexity. You need both sides, and most owners have only one.

Read this table as a strong signal rather than a settled fact. It comes from five businesses in one industry.

One disclosure, since this article is published by a company that sells a check for exactly this. Our free check runs four of these five assistants. Claude costs us roughly double the others to run, so it sits on the paid tier.

Why does ChatGPT mostly quote your own website?

Because it can reach it. Semrush’s clickstream data puts ChatGPT at running a live web search on 34.5% of queries as of February 2026, down from 46% in late 2024, and when the question is about a specific kind of local business, your own pages are the most direct answer available.

That makes your website the biggest single fix for ChatGPT, and it is the one place you control completely. It needs the services you actually offer, written out in words a customer would use, the areas you cover, your hours, and enough detail that a page about water heater replacement says “water heater replacement” somewhere.

Worth keeping in proportion: across the wider research, businesses’ own sites are a small share of what gets cited overall. Seer found brands owned about 2% of the content cited about them. Your website is the lever for one assistant, not the whole job.

Before any of that, spend five minutes checking that your website does not block the crawlers these assistants use. Many sites blocked GPTBot at some point, which is OpenAI’s training crawler, and left the ones that fetch pages during a live answer alone. Some did the reverse by accident. The names to allow are OAI-SearchBot and ChatGPT-User for ChatGPT, PerplexityBot for Perplexity, and Google-Extended if you want to appear in Gemini. Google-Extended does not affect whether you show up in Google Search or in AI Overviews, which run off Google’s normal index.

Check it because it takes five minutes, and treat the evidence with some caution. One study of 1,058 domains found sites blocking OAI-SearchBot were effectively uncited, while a larger analysis by BuzzStream found sites blocking it still appeared in 82.4% of 4 million citations. Nobody has proved the causal link.

Why do Gemini and Perplexity keep citing Reddit?

Because people talk about local businesses there. Reddit was the single most cited source for both in our data.

You cannot fix Reddit, and you should not try. Most subreddits remove self-promotion, and paid mentions tend to get spotted. What you can do is make sure that when someone does name you in a thread, everything a reader finds next confirms you are real and do the work: the same business name everywhere, a website that loads, and reviews with recent dates on them. A recommendation that leads to a dead link is worth nothing.

Is Yelp really the most important one?

For most categories, yes. Across the larger dataset Yelp carries more citations than the other five platforms studied combined, and OpenAI signed a licensing deal with Yelp in July 2026, which points its reviews straight into ChatGPT.

For some categories it does not matter at all.

A note on how we checked this part. The category work below is a different method from everything above. Rather than logging assistant citations, we looked at the search results these categories return in each city. That is a reasonable stand-in for Google AI Overviews, which draw heavily on the same results, and it is not a reliable stand-in for ChatGPT or Perplexity. Treat it as a lead worth checking rather than a measurement, and treat the two-city coverage as thin.

On that basis: personal injury lawyers returned no Yelp results at all in either city we checked. The sites that came up instead were Justia, Super Lawyers, Avvo and Forbes Advisor.

Auto repair was the other surprise. Yelp came first, but Angi and HomeAdvisor, which have been the same company since 2017, did not appear in either city, and neither did Thumbtack. A shop following general home services advice would spend its first afternoon and possibly real money on platforms that never surfaced for its category.

Which websites matter for my trade?

Confidence varies, and we mark it. Where a row rests on the Foundation and AirOps aggregate rather than on our own checks, we say so, because a confidence rating you did not earn is worth nothing.

Two kinds of listing get missed constantly, and both are free. Insurance panels publish public directories, so a dentist already in-network usually has a listing they have never checked. Manufacturers run certified-installer locators, so roofing, HVAC and plumbing contractors who hold a certification usually have one too.

RepairPal for auto repair and RealSelf for med spas would both rank near the top of any list built from reputation. Neither appeared in the cities we checked. We would not publish that as a conclusion yet, and it is the first thing we will look at when the wider runs land.

What about the Better Business Bureau and Nextdoor?

The BBB is worth more than most owners think, and Nextdoor is worth a lot less. Nextdoor came in at 10,300 citations in the Foundation and AirOps data, around 2% of Yelp’s total and about a fourteenth of the BBB’s, against a reputation that suggests far more. If you have limited time, the BBB listing is worth more than the Nextdoor page.

Do I have to pay Angi and Thumbtack?

No. Both platforms blur two different things. A free profile helps you get cited, because the assistants read those pages. Buying leads is a separate purchase with separate economics, and published breakdowns for electricians and HVAC contractors put Thumbtack around $250 per booked job and Angi around $542.

Claim the free profile, and decide on lead buying separately.

What should I do first?

In this order. None of it costs money.

1. Check your robots.txt allows OAI-SearchBot, ChatGPT-User and PerplexityBot, which takes five minutes.
2. Claim and complete your Yelp profile, unless you are in a category above where it does not apply.
3. Verify the listing you already have and have never looked at, whether that is your insurance panel directory, a manufacturer installer locator or your licensing board entry. It exists, and there is a good chance it is wrong.
4. Make sure your own website says, in plain words, what you do and where you do it. This is the fix that reaches ChatGPT.
5. Claim BBB, Angi and Thumbtack profiles without buying leads.
6. Pitch one local editorial “Best Of” list. City magazines, alt-weeklies and tourism boards came up repeatedly across categories and are free to enter.

One thing to add without expecting much, and one to skip. Schema markup, the hidden code some plugins add to your pages, is cheap and standard, so add it, but do not expect it to win you a citation: Ahrefs tracked 1,885 pages that added it against 4,000 that did not and found no meaningful lift on any assistant. An llms.txt file does nothing at all. Of the roughly 38,000 valid ones Ahrefs found across 137,000 monitored domains, 97% received zero requests, and Google’s own guidance says you do not need one.

What we could not measure

Our own logged sample is five businesses in one industry, so the per-assistant ordering is a strong signal rather than a settled fact. The category work is a search-results stand-in across one or two cities each, which is why we label the basis of every row rather than dressing it as one dataset.

A two-city check already caught one error we would otherwise have published: chiropractors returned Yelp first in Boise and no Yelp at all in Charlotte. One city would have given us a confident, wrong answer.

We are running this across more cities and categories, and we will publish what we find.