How to Get Cited by ChatGPT, Perplexity and AI Overviews

The audit you can run in twenty minutes

Do this before reading any further, because the result reframes everything else.

1. Open ChatGPT, Perplexity, Google's AI overview and Claude.
2. Ask each the question your best customer would ask before they knew you existed. Not "tell me about my company". Something like "what are the best tools for X for mid-market companies in Europe" or "who does Y and how much does it cost".
3. Write down which companies are named, in what order, and which sources each answer cites.
4. Repeat with three or four variants of the question.

Two lists come out of that exercise. The companies named are your real competitive set for the next few years, which may differ from the one in your deck. The sources cited are your placement target list, because those are the pages the models are actually reading in your category.

What kind of source gets cited

Across the categories we watch, the sources models reach for share a consistent profile. None of it is exotic, and almost none of it is what classic SEO optimises for.

• Named and dated. A byline and a publication date read as accountable. Undated evergreen pages are weaker, which is the opposite of the old advice to strip dates so content looks fresh.
• Factually dense. Specific numbers, named entities, concrete mechanisms. A model cannot cite an adjective. Marketing prose is skipped because there is nothing in it to quote.
• Published somewhere with a reputation. The domain still matters, for exactly the reason it matters to a human researcher.
• Structured so the answer is extractable. Clear headings, the answer sitting next to the question rather than four scrolls below it, tables, explicit definitions.
• Corroborated. A claim appearing in three independent places carries more weight than the same claim stated once, however prestigiously.
• Indexed and crawlable. If it is not indexed, it does not exist for any engine that matters.

Notice what is absent: keyword density, link attributes, word count, and freshness for its own sake.

The three mistakes that make a company invisible

1. Writing in your own vocabulary instead of the buyer's

If your site says "placement" and your buyer asks about "digital PR", the semantic gap costs you. Models match language. Say the words your category actually uses, including the ones you find imprecise.

2. Publishing claims instead of facts

"Industry-leading accuracy" is unquotable. "We tested 40,000 addresses and the false-positive rate was 0.6%" is a sentence a model can lift verbatim and attribute to you. Proprietary numbers are the most citable asset you own, and most companies sit on them.

3. Never checking whether your sources are crawlable

Some publications block AI crawlers in robots.txt. A placement in a title that disallows GPTBot cannot become a source for ChatGPT, whatever it cost you. Before buying a placement specifically for AI visibility, open the outlet's /robots.txt and look. Almost nobody does this.

What to do on your own site

1. Answer real questions in your headings, phrased the way people ask them, with the answer in the first two sentences underneath.
2. Put numbers in. Your own data, your benchmarks, your survey. Dated and sourced.
3. Use structured data properly. An Organization node with sameAs pointing to your real profiles, articles with author, datePublished and mentions naming the entities you discuss. This is the difference between a machine inferring what your page is about and being told.
4. Be named consistently. Same company name, same spelling, same founder attribution everywhere. Inconsistency fragments the entity and dilutes every mention.
5. Allow the crawlers. Check your own robots.txt permits GPTBot, ClaudeBot, PerplexityBot, Google-Extended and the rest, unless you have a deliberate reason not to.
6. Publish an llms.txt summarising what you do, in plain factual sentences, with links to your key pages. It is not a standard anyone is obliged to read, and it costs an hour.

Turning the audit into a measurement you can report

The twenty-minute audit is useful once. Turned into a monthly habit with four numbers, it becomes the only reporting anyone currently has for this channel.

That last row surprises people and matters more than the rest. If an outdated description of your company, a wrong founder name or a discontinued product is what exists in indexed sources, that is what gets repeated. Correcting the record means publishing accurate, dated, specific material somewhere credible, because you cannot edit a model's memory but you can change what it reads.

Why press coverage is the strongest lever

Everything above is necessary and none of it is sufficient, because a model weighing sources is doing roughly what a careful human researcher does: preferring things published somewhere accountable over things a company says about itself.

That is the whole argument for press coverage in this new layer. A dated, bylined, factually specific article in an indexed publication is precisely the source profile answer engines favour, and it is a source you cannot create by editing your own website. Three of them in independent publications is corroboration, which is the thing that moves a model from mentioning you to naming you.

One further point that surprises people: as far as we can observe, whether a placement carries a commercial label makes no difference to whether it gets cited. What matters is whether the article says something specific enough to be worth quoting. A labelled article full of real numbers is more useful for AI visibility than an unlabelled one full of adjectives, which means the industry has spent a decade optimising the wrong attribute of the same purchase. We go further into that in do media backlinks still work.