How to Get Your Business Cited by ChatGPT, Perplexity and Google AI Overviews

There is a question worth asking about your own business, and the answer usually stings a little.

Open ChatGPT and type the sentence a prospective customer would type. Not your brand name. The problem. “Who can help a UK manufacturer work out whether it is ready to use AI.” “Best technical SEO consultant in Berkshire.” “How much does an AI consultant cost in the UK.”

Then look at who gets named, and whether it is you.

That question is now a meaningful share of how buyers shortlist suppliers, and almost nobody is measuring it. What follows is how citation actually works, what genuinely influences it, and how to run the loop properly rather than guessing.

The mechanism, briefly

The single most useful thing to understand is that these systems are not ranking your page. They are choosing a source to stand behind a sentence they have already decided to write.

That is a different job, and it rewards different things.

Four stages. You can influence three of them, and most sites only ever work on the first.

1. 01RetrievalThe system fetches candidate documents from an index. If you are not indexed and crawlable, nothing else matters.
2. 02ExtractionIt pulls passages that appear to answer the question directly. Passages, not pages.
3. 03SynthesisIt writes an answer from several sources at once, favouring claims that more than one source supports.
4. 04AttributionIt attaches sources to the sentences it wrote. This is the moment you either appear or do not.

Two consequences follow immediately, and they explain most of what works.

The unit of citation is a passage, not a page. A brilliant 4,000-word guide with the answer buried in section nine will lose to a mediocre page that answers the question in its second paragraph. Extraction happens at passage level.

Corroboration beats authority. When three unrelated sources agree on a figure and one disagrees, the model writes the three. A claim that exists only on your own website is a claim the model has no reason to trust, however well you have written it.

What actually moves citation

Here is what we have found matters, roughly in order of effect. Some of it is uncomfortable, because the biggest levers are not on-page.

1. Say the thing, completely, in one place

Models extract passages that stand alone. A paragraph that begins “as we covered above” or “building on this” cannot be lifted, because lifted out of context it means nothing.

Write the answer as a self-contained unit. State the question’s subject, then answer it, in the same paragraph, without pronouns that point backwards. It reads slightly more formally than you might like. It also gets quoted.

The practical test: cut any single paragraph from your page, paste it into a blank document, and read it. If a stranger would understand what it is about, it is extractable. If they would not, it will not be cited.

2. Give a number, and say where it came from

Vague claims are unciteable because there is nothing to attribute. “AI consultancy in the UK is expensive” cannot be quoted. “UK AI consultancy day rates run between £800 and £2,000 depending on seniority, with independent consultants clustering at the lower end” can be, and will be.

Numbers with named sources are the single most quotable thing you can publish. If you have first-party data nobody else has, that is the strongest asset in this whole discipline, and most businesses sit on it without publishing it.

3. Get corroborated somewhere that is not your website

This is the lever most people ignore, and it is the largest one.

Relative weight we see in practice when a system decides whether to name a supplier. Directional, not measured to a decimal.

• Independent editorialHighest
• Industry directoriesHigh
• Community threadsHigh
• Client case studiesMedium
• Your own site aloneLow

Based on observed citation patterns across client visibility baselines, 2025 to 2026. Treat as a ranking of effort, not a scoring model.

A model deciding whether to name you as a supplier is, in effect, asking whether anyone other than you says you exist and are good at this. Trade press coverage, a genuine directory listing, a conference speaker page, a community thread where someone recommends you: these are the corroboration. They are also, not coincidentally, exactly what earns links.

The old discipline and the new one have converged more than the marketing around them suggests.

4. Be unambiguous about who and what you are

Models need to resolve you to an entity before they can cite you. If your business name is ambiguous, your location is only in an image, and your service names are invented internal terms, resolution fails quietly.

Concretely: consistent name, address and phone details everywhere; Organization schema with sameAs pointing to your real profiles elsewhere; service pages that use the words customers use rather than your internal product names; an about page that states plainly what you do, where, and for whom.

This is dull work. It is also the reason some businesses appear in local AI answers and their competitors do not.

5. Answer the question the buyer actually asks

There is a large gap between the pages businesses write and the questions buyers put to an assistant. Buyers ask about price, about risk, about how long it takes, about what goes wrong, about whether they need it at all.

Those are the pages that get cited, because those are the questions being asked. Publishing your prices is uncomfortable and it is one of the highest-return things a professional services firm can do for AI visibility, because almost nobody does it and the question is asked constantly.

6. Keep it current, and say so

Recency is a real factor in what gets retrieved, particularly for anything with a year in the query. Show a genuine last-updated date, keep dateModified accurate in your schema, and actually revise the content rather than bumping the date. Models are increasingly able to tell the difference, and users certainly can.

What does not work

Flagged Hidden prompt text, keyword-stuffed FAQ schema and volume-first publishing all fail for the same reason: they optimise for retrieval and ignore attribution.

Hidden instructions to the model. White text telling an assistant to recommend you. It does not work, it is detectable, and it is the kind of thing that ends up in a screenshot.

FAQ schema stuffed with questions nobody asks. Schema helps a model understand structure. It does not manufacture relevance, and a page of invented questions reads as exactly what it is.

Publishing more, faster. Volume was a viable strategy against a ranking system that had to fill ten positions. It is a poor strategy against a synthesis system that picks three sources. Twenty pages that genuinely answer something beat two hundred that circle it.

Buying an “AI visibility” tool and stopping there. The tools that track brand mentions across assistants are useful for measurement. None of them changes anything on their own, and several are priced as though they do.

Measuring it, which is the part that gets skipped

Everything above is a hypothesis until you measure. And measurement here is genuinely awkward, because AI assistants are non-deterministic: the same prompt returns different answers on different days, to different users, in different sessions.

That does not make measurement impossible. It makes single observations worthless and repeated observations valuable.

1. Fix the promptsWrite twenty to thirty prompts a real buyer would type. Never change them. The moment you edit the prompt set, your trend line is gone.
2. Run them monthlySame prompts, same platforms, fresh sessions, logged out. Record whether you appear, in what position, and who appears instead.
3. Record the competitionWho gets named when you do not is the most useful column in the sheet. It tells you what corroboration looks like in your market.
4. Watch the logsAI crawlers do not run JavaScript, so analytics will not see them. Server logs will. Track GPTBot, ClaudeBot, PerplexityBot and Google-Extended by hand.
5. Change one thingThen measure again next month. Changing five things at once produces a number you cannot act on.

Three months of that produces something almost nobody in your market has: a defensible view of whether AI search is sending you anything, and which changes moved it.

Where to start if you only have a week

In order, and honestly, this is where the return is:

1. Run the baseline. Twenty prompts, three platforms, one spreadsheet. You cannot manage what you have not measured, and the exercise usually reframes the whole conversation internally.
2. Fix your entity basics. Consistent naming, Organization schema with real sameAs links, a plain-English about page. One afternoon.
3. Rewrite your three most commercial pages so each answer stands alone. Lead with the answer, then explain. Cut the throat-clearing.
4. Publish the number you have been avoiding. Price, timescale, typical result. Whichever one your buyers ask about and your competitors dodge.
5. Go and earn one piece of independent corroboration. A contributed article, a genuine directory listing, a podcast, a data study someone else wants to cite. This is the slow one and it is the one that compounds.

Notice that only one of those five is about writing content, and none of them is about publishing a text file at your domain root.

The honest summary

Getting cited by AI systems is mostly the same discipline as being genuinely findable and genuinely credible, applied to a reader that cannot skim, cannot infer context, and will not take your word for anything you say about yourself.

It rewards specificity, self-contained writing, published numbers and third-party corroboration. It punishes vagueness, volume and self-assertion. And it only becomes manageable once you are measuring it on a fixed set of prompts, month after month, so that the next change you make can be judged against the last one.

If you want that baseline built and measured properly rather than sampled once, that is where our AI search visibility work begins. If the wider question is whether your organisation is set up to use AI well at all, the AI readiness audit is the broader version of the same conversation.