No one can promise a local business an AI mention, but the evidence the assistants read can be measured

Three assistants, one question, three different answers

I asked ChatGPT, Gemini and Perplexity the questions a patient in Bradford might ask: which clinics offer a treatment, who is well regarded, what my own clinic does and where it is. The three did not agree with each other, did not always agree with themselves when asked again, and some of what they said was plainly wrong.

When I traced the wrong statements, the trail did not end at the assistant. It ended at third-party pages: a directory listing nobody had updated, a profile that described the clinic differently from the website, pages that contradicted each other. The assistant had not invented the error; it had repeated it.

That is the frame for this piece: what is actually documented, by the platforms and by studies that can be opened in full, on how these systems choose which local businesses to name, and how much of that a business can influence. The answer is less than the GEO industry implies and more than a sceptic might assume.

I am a prescribing pharmacist and the founder of Pulse & Clarity Clinic in Bradford. I built the clinic's website, Google Business Profile, analytics and local search presence myself during 2026, without an agency, and the tests described here included my own clinic.

The scale is no longer small. Ofcom reports that ChatGPT had 1.8 billion UK visits in the first eight months of 2025, up from 368 million a year earlier, and that about 30 per cent of Google searches now show AI Overviews ^[15]. Those figures describe use in general; that the growth reaches a person choosing a plumber is a reasonable inference, not a measured fact.

What the platforms say they retrieve

The most useful documents are the ones the platforms wrote about themselves.

Google

Google's guide to its generative AI features describes retrieval-augmented generation, which it also calls grounding, as improving AI responses "by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index", and query fan-out as concurrent, related queries generated by the model to fetch additional results ^[2]. Its documentation for site owners states that there are "no additional requirements to appear in AI Overviews or AI Mode" and no special optimisation is necessary ^[1].

For local businesses, the guide says that using Google Business Profiles "can help your products and services to be visible in both AI responses and other Google Search results", and that the models use publicly accessible, crawlable content ^[2].

ChatGPT

OpenAI's help documentation says ChatGPT may search the web automatically when a question would benefit from current information, and that it "typically rewrites your query into one or more targeted queries that it sends those providers" ^[4]. The providers are third parties; the launch announcement said ChatGPT search draws on "third-party search providers, as well as content provided directly by our partners" ^[5]. For local results, ChatGPT "may use an approximate location based on your IP address to provide relevant local results", with precise device location optional and off by default ^[4].

OpenAI's crawler documentation states that sites opted out of its OAI-SearchBot crawler "will not be shown in ChatGPT search answers, though can still appear as navigational links", and separates that crawler from GPTBot, which is used for training ^[6].

Perplexity

Perplexity describes itself as searching the internet and gathering information from articles, websites and journals, with each answer including "numbered citations linking to the original sources" ^[7]. Its crawler documentation says PerplexityBot "is designed to surface and link websites in search results on Perplexity" and is not used to train models ^[8].

Gemini

Google's developer documentation for grounding Gemini with Google Search says the model "analyzes the prompt and determines if a Google Search can improve the answer", then runs searches and produces a response grounded in the results, with citations ^[9]. In the consumer Gemini app, Google's help page notes that "Not all responses include related links or sources" ^[10].

Read together, these documents establish one thing firmly. The local business is found the way a search engine finds it, by retrieving pages from an index and summarising them; there is no registry inside the model that a business can apply to join. That makes Google's local ranking page relevant to AI answers too. It says local results are based primarily on relevance, distance and prominence, that businesses with complete and accurate information are more likely to appear, and that "There's no way to request or pay for a better local ranking on Google" ^[3].

How much the answers change between runs

The strongest published evidence on variability comes from SparkToro, whose founder Rand Fishkin ran a volunteer study in November and December 2025 in which 600 people submitted 2,961 runs of 12 prompts through ChatGPT, Claude and Google's AI Overviews and AI Mode, copying each response into a survey form ^[11].

There's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses.

On ordering the picture was worse: the study found it was closer to one in 1,000 runs before two lists would appear in the same order, and that nearly every response differed in the list, the order and the number of items ^[11].

Two caveats. The prompts asked for brands and products rather than a local clinic or tradesperson, so the numbers cannot simply be transferred to a query with a town name attached, and volunteer copying is a coarser instrument than an automated harness. Neither rescues the idea that a single screenshot tells a business anything reliable.

Variability shows up inside Google's own features as well. Ahrefs compared 863,000 keyword results with 4 million AI Overview URLs collected in January 2026 and found that "just 38% of AI Overview citations come from top 10 pages, down from 76% a year ago", with the rest split almost evenly between pages ranked 11 to 100 and pages outside the top 100 ^[14]. The authors attribute the change to query fan-out, with citations arriving from related sub-queries rather than the original search ^[14]. This is a study of AI Overviews across keywords generally, not of local packs, but it fits the mechanism Google describes, and it means that ranking well for a phrase does not make a business certain to be cited for it.

Which sources get cited for local questions

The largest open dataset on local citations comes from BrightLocal, which in August 2026 published an analysis of 60,970 checks across 1,355 business locations in the United States, United Kingdom and Australia, running five prompts at nine map points per location against Google AI Overviews, AI Mode and ChatGPT, and tracking 1,967,762 citations across 116,670 domains ^[12].

Google Business Profile accounted for 28.5 per cent of all citations, and Yelp for 8.53 per cent, appearing as a source in 80 per cent of ChatGPT's local answers ^[12]. The UK directory Yell appeared in the top 50 with 709 citations across 68 businesses ^[12]. The finding that matters most for a small business is a different one: around 93 per cent of the domains cited were individual business websites ^[12].

This is a vendor study using the vendor's own tracker, and the three countries are pooled, so the weight of Yelp says more about American answers than about a Yorkshire dentist. The pattern is nonetheless consistent with the platform documents.

For Google's features there is a first-party measurement. Search Console counts a click on a link inside an AI Overview or AI Mode as a click and assigns every link in an AI Overview the single position the Overview occupies ^[16], and Google's guide points site owners to a generative AI performance report there ^[2]. Nothing comparable exists for ChatGPT, Gemini or Perplexity; a business wanting to know what they say has to ask them.

The GEO industry, and what it rests on

Much of the advice sold as generative engine optimisation descends from a single academic paper. In November 2023 researchers at Princeton and elsewhere published the paper that coined the term GEO, later presented at KDD 2024, which tested nine ways of rewriting a web page to see which increased its share of a generated answer ^[13]. The paper deserves its due. It found that the best methods improved on the baseline by 41 per cent on its position-adjusted word count metric and 28 per cent on subjective impression, that adding quotations and statistics produced relative improvements of 30 to 40 per cent, and that keyword stuffing offered "little to no improvement" ^[13].

Its limits matter just as much. The engine in the experiment was built by the authors: it fetched only the top five Google results for each query and passed them to the gpt-3.5-turbo model to write the answer ^[13]. The queries were general, not requests for a local business, and the authors note that "the efficacy of these strategies varies across domains" ^[13]. What the paper shows is that a page giving a model something concrete to quote does better, in a simulated engine, than a page full of adjectives. That is sensible. It is not a recipe for being named the best dentist in Halifax.

The llms.txt file

The most visible of the newer proposals is llms.txt, a text file meant to tell language models what a site is about. In April 2025 Google's John Mueller wrote that, as far as he knew, none of the AI services had said they were using it and that server logs showed the bots did not check for it, comparing it to the keywords meta tag ^[17]. Google's guide, last updated in July 2026, is explicit: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search" ^[2]. Search Engine Journal reported in May 2026 that Google's position differed between products, with the Search team saying the file should be skipped ^[18]. No platform document fetched for this piece describes reading it.

Structured data and chunking

Google's guide says that "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add", and that there is "no requirement to break your content into tiny pieces for AI to better understand it" ^[2]. Schema still describes a page accurately to a crawler; it does not open a separate door.

A good deal of GEO advice is ordinary good practice under a new name: clear pages, real information, accurate listings. Google's guide makes the same point, describing work on generative AI search as being about the search experience "and thus still SEO" ^[2]. The problem is not the advice but the certainty attached to it.

What a business actually controls

If the assistants retrieve and summarise, what a business controls is the material available to be retrieved. Three properties of it can be worked on and checked.

Consistency

Google's local ranking page says businesses with complete and accurate information are more likely to show up in local results ^[3]. The wrong statements I traced were not hallucinations in the usual sense; they were faithful summaries of stale or contradictory pages, and the problem sat on the directories and profiles, not on my website. Name, address, phone number, services, practitioners and hours should read the same everywhere a crawler looks, and the crawler must be allowed in: Google requires crawlable content ^[2], OpenAI recommends allowing OAI-SearchBot ^[6] and Perplexity recommends allowing PerplexityBot ^[8].

Depth

Google's guide distinguishes content people find unique, compelling and useful from what it calls commodity content, which "could originate from anyone, and typically adds little unique insight for readers" ^[2]. For a local business the equivalent is not an essay but a page that states plainly what is offered, by whom, with what qualifications, at what guide price, where and when. A model can only repeat what is stated somewhere; ten pages about excellence and passion give it nothing to state.

Spread

The BrightLocal data suggests the assistants read the Google profile first, directories and review platforms second, and the business's own site throughout ^[12]. Which directories matter varies by sector and country, and the way to find out is to run the questions and record what is cited in that market; presence and accuracy on the sources that actually appear is worth more than presence on hundreds that do not. Google says more reviews and positive ratings can help local ranking ^[3], and the same profile is the single most cited source in local AI answers ^[12]. None of that justifies offering anything for a review; it justifies asking properly.

How to test without fooling yourself

Because two runs of the same prompt rarely produce the same list ^[11], a test is only meaningful when repeated and only comparable when the prompts are fixed. SparkToro's advice is to track a visibility percentage across dozens to hundreds of prompts run multiple times, and to treat any tool claiming to report a ranking position in AI as unreliable ^[11].

A workable method looks like this.

1. Write a fixed bank of realistic questions in a customer's words: who is good for a service in the town, who to recommend, what people say about a named business, what it costs.
2. Run the same bank on each platform, several times, on a schedule, and keep every response.
3. Record for each response whether the business is mentioned, whether its website or profile is cited, and whether what is said about it is accurate.
4. Record which sources were cited for every competitor named; that list is the map of where evidence about the market is read from.
5. Compare the mention rate and the citation list month on month, and ignore any single run.

The three recorded properties are distinct achievements: being mentioned, being cited as a source, and being represented accurately. The third is the one most businesses discover last, usually when an assistant tells a prospective patient that a treatment is not offered when a room has been fitted out for it.

HighRegard publishes a free Live AI Visibility Test that runs three prompts, covering best in town, a recommendation and reputation, against one commercial AI model with live web search, and records whether the business is mentioned, how many competitors are named and which sources were cited ^[19]. Its page describes the result as a snapshot rather than a study, notes that the same prompt can produce different answers an hour later, and says proper measurement means running a bank of prompts across providers on a schedule ^[19]. I built it because the first thing a business needs is to see what an assistant actually says and which sources it read; the second is to stop trusting any one answer.

What this means for a local business

The practical conclusions follow from the documents rather than the sales material.

• There is no submission form, file or tag that puts a business into AI answers; Google says so directly ^{[1] [2]} and the other platforms describe ordinary crawling and search ^{[6] [8]}.
• The Google Business Profile is both the first input to Google's local results ^[3] and the most cited single source in local AI answers ^[12], so it should be complete, accurate and kept that way.
• The business website is read directly and at scale ^[12]; it should state services, people, credentials, prices, location and hours in plain text a model can quote.
• Third-party pages that describe the business differently are a liability, because the assistant may repeat them; correcting them is measurable work.
• Any claim that a supplier can secure a recommendation from ChatGPT or Gemini should be tested against the platforms' statements that nothing beyond ordinary, crawlable, useful content is required ^{[1] [2]}.
• Testing should be repeated and recorded; a single favourable or unfavourable answer proves nothing in either direction ^[11].

For a regulated business such as a clinic, the accuracy test matters more than the mention test; a wrong statement about which services are offered, or by whom, is a problem in its own right.

What the evidence does not settle

First, none of the platform documents explains how the model chooses among retrieved pages when composing a list of businesses. Google describes retrieval through its ranking systems ^[2] and OpenAI describes queries sent to third-party providers ^[4], but the step from ten retrieved pages to three named businesses is not documented anywhere fetched for this piece.

Second, the variability figures come from brand and product prompts ^[11] and from AI Overviews across general keywords ^[14], not from local service queries. Local answers might be more stable, because the candidate set in a town is smaller, or less stable, because the evidence about small businesses is thinner. Nothing fetched here measures it.

Third, the citation data is pooled across three countries and produced with a vendor's own tracker, with no UK-only breakdown by sector ^[12].

Fourth, the academic evidence that content changes increase a page's share of an answer comes from a simulated engine using a 2023 model and five results per query ^[13]; whether the effects hold in live systems that fan out queries across far more sources ^{[2] [14]} has not been shown.

Fifth, nothing above measures whether being named produces a booking. The HighRegard tool's page says it cannot measure whether a mention converts into customers ^[19], and no source fetched here does either.

What remains uncertain

What the evidence leaves a business is unglamorous and checkable: make the evidence about itself consistent, deep and present where it is read, then measure the mention rate and the accuracy of what is said, repeatedly, over months. Whether that work moves the mention rate, by how much and how quickly, is the part no one has yet shown in a way that can be opened and read. I have traced wrong statements about my own clinic to the pages they came from. I have not seen, and would not claim, that anything guarantees a name in an answer.

The honest position is that the assistants are a new reader of the same evidence, and the evidence, unlike the reader, is something a business can change.