How ChatGPT, Perplexity and Google AI Overviews choose which businesses to cite

An AI answer can name your business without citing your website. It can also cite your website for a useful explanation without recommending your business. These are different outcomes. Before trying to improve visibility, decide whether you want to be mentioned, used as a source, or considered by someone choosing a supplier.

There are understandable reasons why a source might be useful: it answers the question, identifies who is speaking, supplies checkable evidence and can be read. But the public documentation does not give us a complete citation formula. Below, documented behaviour means the provider describes it. Practitioner inference means a reasoned explanation, not a disclosed ranking signal.

Separate discovery from citation

For a web-grounded answer, distinguish three jobs: finding candidate material, using relevant information to compose an answer, and attaching supporting links. That is a useful conceptual model, not a claim that all three products implement the same pipeline. Passing one stage does not prove you will pass the next.

A crawler visit establishes access, not selection. A retrieved page might contribute nothing to the final response. A citation might support only one sentence. Equally, a business recommendation might draw on a directory or review instead of the business's own site. Counting these together hides what actually happened.

The engines have different starting points

ChatGPT can answer without searching, so this comparison concerns its search features. OpenAI's crawler documentation distinguishes OAI-SearchBot, which supports search visibility, from GPTBot, which concerns potential training use. It explicitly allows separate permissions. Letting a training crawler in is not a prerequisite for search inclusion.

Perplexity describes searching the web, summarising information and linking to sources. It also offers different search modes. Its crawler documentation recommends allowing PerplexityBot and its published IP ranges. Those are access instructions, not a promise of selection.

Google's AI features guidance says a supporting page must be indexed and eligible for a Search snippet. It also documents possible query fan-out: related searches across subtopics and sources. An Overview need not use only results for the exact words typed, and Google does not show one for every query.

These differences matter when diagnosing absence. Google indexing is a stated condition for Overviews; it is not a universal inclusion test for the other products. None of the guidance above supplies weights for comparing two businesses or lets us calculate a citation probability.

Answer the question the buyer actually asked

Start from the information needed to answer. If someone asks whether a supplier serves a particular area, a clear service-area statement provides evidence. A paragraph calling the supplier innovative does not. If they ask about cost, explain the price, scope, exclusions and conditions together. A number without its conditions can mislead.

Our practitioner inference is that explicit, self-contained explanations make a page easier to use accurately. The reasoning is simple: the answer requires certain facts; a passage containing those facts demands fewer unsupported assumptions. That does not establish a preference for a particular paragraph length, keyword density or question-shaped heading.

Put the answer near its qualifications. Explain when a service is unsuitable and what the customer must supply. Use a table for a real comparison and a list for steps. Do not manufacture dozens of near-identical questions to suggest depth that the business cannot substantiate.

Consider a hypothetical supplier comparison. If the question requires installation, maintenance and coverage in Leeds, evidence of maintenance alone does not establish the full fit. A useful page states each capability and its limits. This is reasoning from the question, not a finding about any engine: leaving those details unstated asks the answer writer to fill gaps that you could have resolved.

Make the business unambiguous

Entity clarity means making it clear which organisation, person or place a statement concerns. A trading name, legal company name and branch name can all be correct, but their relationship needs explaining. State who provides the service, where they operate and which contact details belong to them.

Consistency across your site and relevant external profiles is sensible information hygiene. Our inference is that fewer contradictions reduce the work needed to associate facts with the right business. We cannot turn that into a documented universal entity score. Nor should consistency mean claiming every branch has identical opening hours or services when it does not.

Make corrections where the contradiction lives. If an old directory lists a former address, updating your homepage leaves that conflicting source available. If a brand changed names, explain the change rather than silently replacing every reference. The practical objective is that a reader encountering either version can establish the relationship, without relying on an assumed machine reconciliation process.

Structured data helps describe, not prove

Google documents using structured data to understand page content and information about things such as companies. Markup can explicitly label a business name, address or relationship. Its useful role is description. It cannot independently verify that your claims are true.

Google's AI features guidance says no special schema or new AI text file is required. Markup should agree with visible content. We would not extrapolate Google's documented use into a claim that ChatGPT or Perplexity awards a citation boost for a particular schema type. Validate accurate markup, but fix an unclear service page before adding more labels to it.

Independent references can add evidence

Google's ranking systems guide documents link analysis, including PageRank, within Search. That supports the relevance of links to search discovery and ranking. It does not disclose a direct conversion from backlinks into AI Overview citations, much less a shared rule for all three engines.

If an engine retrieves a relevant trade publication that independently describes your work, it has evidence beyond your own assertions. Our practitioner inference is that references from sources the system considers useful may help it discover or corroborate a business. We cannot inspect a universal trusted-source list or promise that trust passes automatically through a link.

The reference must support the claim. A membership listing can support membership, not exceptional service. Coverage based on your press release is not independent verification of every sentence. Seek accurate coverage that gives a reader something to check, rather than treating every mention as an interchangeable authority token.

Recency matters when facts can expire

The same Google ranking guide documents freshness systems for queries where recent information is expected. That is narrower than saying new pages always win. For a current price or product specification, an old answer may be wrong. For an explanation of a stable concept, age alone does not invalidate it.

For the other engines, treating current evidence as useful on time-sensitive questions is a practitioner inference here, not a claimed freshness weighting. Update changed facts, state when information applies and retain useful context. Changing a publication date without reviewing the underlying content supplies no better evidence.

Make the evidence machine-readable

Check the response a crawler can receive, not just the page in your browser. A firewall challenge, login or failed request can prevent access even when robots.txt permits it. Inspect server logs and verify legitimate crawler identities against provider guidance before changing security rules. Permission and successful delivery are separate checks.

Publish essential copy as readable HTML, with meaningful headings, real links and labelled tables. Test with JavaScript disabled. That is a robustness recommendation: it removes dependence on a visitor executing scripts to obtain the facts. It is not a claim that all engines lack rendering capabilities or award a semantic-HTML citation bonus.

When reviewing an answer, open its citations. Ask whether each source supports the nearby statement, whether your business is represented accurately and whether the question has commercial relevance. Keep the prompts stable enough to compare, and retain unsuccessful checks alongside successful ones. A screenshot of one favourable answer is not evidence that buyers routinely see it or that it produces enquiries.

What a business can actually do

1. Choose real buyer questions and identify the facts needed to answer each one.
2. Check access, then correct missing answers, contradictory business details and expired information.
3. Publish evidence you can stand behind and seek relevant independent references.
4. Repeat documented checks across engines, then assess enquiries as well as visibility.

WebWisp's programmes cover question research, technical foundations, content and measurement. You still need to supply accurate business facts and evidence. We can improve what is available to be found and understood; the engines control selection. Our how it works page explains the delivery process.