What AI Assistants Answer When Asked Who to Hire: 141 Answers Analysed

Most writing about AI search optimisation describes a mechanism without measuring it. This is a measurement: 47 buying questions, three assistants, 141 answers, every citation counted.

The findings that follow are about mechanisms rather than percentages, because 47 prompts is enough to see how these systems behave and not enough to put confidence intervals on anything. Where a number is soft, it says so.

What is generative engine optimization, and what this study measures

Generative engine optimization, usually shortened to GEO, is the work of getting a business named inside an answer written by an AI assistant rather than inside a list of blue links. Adjacent terms — LLM SEO, answer engine optimization, AI visibility — describe the same job from different angles.

The question this study asks is narrower and more testable than "how do I do GEO". It is: when a buyer describes their situation to an assistant and asks who to hire, which providers get named, and where does the assistant get those names from?

That second half is the part almost nobody publishes, and it turns out to matter more than the first.

How this generative engine optimization study was run

Forty-seven prompts, written the way buyers talk to assistants rather than the way anyone types into a search box. Not "web development Warsaw" but "I run a small law firm in Warsaw and need a multilingual website that also ranks in Google. Which developer or small agency should I contact?"

The prompt set covered eight blocks:

Each prompt asked for names explicitly. Without that instruction an assistant tends to return advice instead of a shortlist, which measures nothing.

Three engines, every prompt through each:

Total run cost: about 1.50 USD. The method is reproducible for any domain and is written up step by step in how to check what AI assistants say about your company.

How much each assistant cites per answer

Assistants do not return ten links. They return a shortlist, and how much evidence sits behind that shortlist varies by an order of magnitude.

The practical consequence is a change of shape, not of degree. A results page has a first page, a second and a third. An answer has room for a handful of names. There is no page two: a business is either in the shortlist or it does not exist for that question.

The spread also sets expectations for how much a website can influence each engine. Perplexity reads widely enough that a well-written page has room to be found. ChatGPT, at 1.6 citations per answer, is selecting from a much narrower set, and the selection happens before the page is read.

Where each assistant takes its sources: ChatGPT, Perplexity and Google AI Mode compared

This is the finding with the most practical weight, and it is the reason "optimising for AI search" is not one job.

Every citation across all 141 answers was classified into four buckets: the provider's own website, a directory such as Clutch or Sortlist, a platform such as LinkedIn or Reddit, and Google Maps.

Perplexity cited provider websites 667 times, roughly fourteen per answer. It reads the open web and shows its work.

ChatGPT cited provider websites 47 times across 47 answers — almost exactly one per answer — and referenced Google Maps in half of them.

Google AI Mode cited least of all, around three sources per answer, with the tightest selection.

Three engines, three different reading habits. Work that makes Perplexity cite a business is work on that business's website. Work that makes ChatGPT name it is largely not on the website. Any single service sold as covering "AI search" is averaging three different problems.

Why ChatGPT's Google Maps citations are not business listings

Half of ChatGPT's answers referencing Google Maps invites an obvious conclusion: a verified Google Business Profile must be the entry ticket for AI visibility. The link format disproves it.

Every Maps link in the dataset is a search query the model composed from a name and a city. Not one points at an actual listing.

ChatGPT finds providers through ordinary web search and then renders a Maps search link beside each name as a convenience for the reader. Whether the business has a verified profile does not enter into the selection.

A Google Business Profile remains worth having for the local pack and for map searches performed by people. It is not the mechanism that gets a business into an AI answer, and advice to the contrary can be checked against the shape of the URL in about a minute.

LLM SEO versus classic SEO: the two channels reward different pages

Asked about generative engine optimization as a subject, assistants cite these sources:

The Google results page for the same commercial terms looks nothing like that. There, the first page is roughly half agency landing pages and half listicles: "The 8 Best GEO Agencies for B2B SaaS Brands", "10 Best Generative Engine Optimization Agencies", "Best LLM SEO Agency: We Reviewed 22 for AI Visibility".

Two channels, two different sets of winners, and effort spent on one does not transfer to the other. This is the most expensive misunderstanding in the category.

What kind of provider gets named for generative engine optimization services

Individual company names are withheld: this is a study of a mechanism, not a ranking of firms, and publishing a shortlist built from 47 prompts would overstate what the sample supports. The pattern is the finding.

Three things in that table are worth more than any list of names.

A dedicated page beats a section. In the UK agency set, the providers named overwhelmingly had a page whose entire subject was the service being asked about. Providers offering the same service as a section inside a broader SEO page were largely absent, even where the broader page was stronger overall.

One provider was cited through a directory profile rather than its own website. Its homepage did not make the answer; its listing in an agency directory did. If there is a single repeatable tactic in this study, that is it.

For the developer prompts, assistants named people rather than companies. Each had a personal site, and in one case a page whose URL was literally the service being asked for. In this niche being a single specialist is not the disadvantage it looks like. Being unreadable is.

The category boundary moved, too. For the entity-maintenance prompts, two of the five named providers were reputation-management and Wikipedia specialists rather than search agencies. An assistant asked a question phrased around sources and records does not restrict itself to the industry that usually claims the keyword.

What makes an assistant name one provider over another

In the clearest case in the dataset, Perplexity explained its own choice. The provider name is redacted; the reasoning is quoted exactly:

“

For a small law firm in Warsaw that needs a **multilingual website with

“

SEO, the strongest fit from the results is [provider]**: they

“

explicitly say they build websites and run SEO for Warsaw businesses,

“

work in Polish, English, and Russian, and offer a multilingual,

“

conversion-focused site with search/AI optimisation **included in their

“

package descriptions**.

The model matched four stated facts against four conditions in the question: what the work is, who it is for, which languages, what is included. It did not assess quality, weigh a portfolio or read reviews. The operative phrase is "they explicitly say".

That produces a rule worth stating flatly. An assistant recommends the business whose specialisation a machine can read without inference. Claims that require a human to interpret them are invisible to this process.

The consideration set and the recommendation are different states

Four further answers in the dataset cited a provider's page as a source and then recommended somebody else. The page was good enough to inform the answer and not good enough to win the slot.

Most advice about AI visibility does not distinguish these two states, which makes it hard to diagnose. Being read is necessary and not sufficient.

Entity ambiguity outranks page quality

One pattern in the data explains more failures than any on-page factor. Where a provider's name is shared with other people or companies, assistants name a competitor whose identity is unambiguous instead.

A brand-name search that returns a mix of unrelated people, a company in another industry and a namesake in the news leaves a machine unable to resolve which entity the question is about. No amount of rewriting service pages fixes that. The fix is entity work: one canonical name, a Person or Organization node with a stable identifier, every profile listed under sameAs, and the same description word for word across the site, the directories and the social profiles.

AI visibility tracking: how to measure this for your own business

The run above is repeatable at a cost that makes monthly tracking trivial.

The metric that matters is the count of answers naming the business, tracked against a fixed baseline. Share-of-voice percentages across a small prompt set move for reasons that have nothing to do with the business.

Limits of this study

Assistant configuration is not user configuration. The ChatGPT results come from gpt-4.1-mini via API with web search. The consumer application uses a different model, a different retrieval stack, personalisation and memory. Results from one do not transfer to the other.

One snapshot. Answers vary by phrasing, language, session and time, and are reassembled continuously. Nothing here is a stable ranking.

Sample size. Forty-seven prompts show the shape of the mechanisms. They do not support percentage claims, and none are made.

Language coverage. Nine of the 47 prompts were in Russian or Polish. The English findings rest on a larger base than the other two.

AI visibility checklist: what this data actually supports

1. State what the business does, for whom, in which languages, and what is

included, in plain declarative sentences. The one answer that explained its own choice quoted exactly that.

1. Resolve entity ambiguity before anything else. If a search engine cannot

tell the business apart from its namesakes, the rest does not matter.

1. Choose which assistant matters for the buyer in question, because optimising

for one is not optimising for another.

1. Get into the directories that are actually cited, not merely the ones that

accept registrations.

1. Treat video and professional networks as source material rather than

distribution, because that is where the citations come from for this topic.

1. Measure monthly against a frozen prompt set.