How to get your content cited by ChatGPT and AI search

Generative engine optimization, or GEO, is the work of getting your content quoted by AI answer engines: ChatGPT, Perplexity, Google's AI Overviews, Gemini, Copilot. It shares foundations with SEO and chases a completely different event. SEO wants a ranked link that earns a click. GEO wants a sentence of yours to show up inside somebody else's answer, with your name attached.

Almost all of the work is editorial. You can do it this week without filing a single engineering ticket.

You can also do most of it wrong, because this field has more confident advice than evidence. I have repeated some of that advice myself, in an earlier draft of this article, which is a large part of why the corrections below are here.

Three levers, and one that does nothing

The levers are about writing, not infrastructure. The one peer-reviewed study anybody actually cites here is Aggarwal et al., published at KDD 2024, and it tested a pile of content modifications to see which ones made generative engines quote a page more. Three won: cite your sources, add quotations, add statistics. Keyword stuffing, the reflex we all spent fifteen years building, produced what the authors politely call "little to no improvement".

Fifteen years. Little to no improvement.

Write so a paragraph survives being ripped out of context

The core skill in GEO is writing passages that stand on their own. An engine does not quote your article. It quotes a fragment, stripped of the heading above it and the example below it, and hands that fragment to somebody who will never see the rest. Any sentence whose meaning depends on the missing material is invisible, no matter how well it reads in place.

Your best paragraph might be unquotable. That is a hard thing to accept about your own writing.

• Name the subject in the sentence instead of leaning on a pronoun that points three paragraphs back.
• Put the answer in the first sentence under each heading, then support it. Never build to a conclusion.
• One claim per paragraph, so a retriever does not have to guess which half you meant.
• Keep the number and its source in the same sentence. A statistic separated from its citation is one an engine is right to distrust.
• Cut openings like 'as we saw above'. They explicitly mark a passage as unliftable.

Sounding confident buys you nothing

Turning up the persuasion does not get you cited. The GEO authors tested exactly that. They rewrote content in a more authoritative, confident register, measured it, and found no significant change. The engines did not care.

You will nonetheless find it stated everywhere that promotional writing carries a 26% citation penalty. That number does not come from the paper. It traces back to vendor correlation studies, and I could not find a primary source for it anywhere. I had it in this article for a week.

So the honest version is duller than the myth and considerably more useful. Rhetoric is neutral. Evidence is the lever. Every hour spent making a paragraph sound more authoritative is an hour not spent giving it a number and a source, and only one of those two things measured.

Write like documentation anyway. Say what the thing does, state the limits, name the cases where it is the wrong choice. Not because superlatives are punished. Because a page full of them almost never has a fact in it, and facts are the part that gets quoted.

Admitting a boundary is also the cheapest credibility available. It costs one sentence.

Schema: useful, and not what you were told

Google says plainly that you need no special markup to appear in its AI features. The exact wording in its own documentation is that there is "no special schema.org structured data that you need to add". Read that twice, then reread the last five posts you saw promising a specific percentage lift from schema. Every one of them is selling a vendor correlation.

Ship it anyway, for the two things it genuinely does: it makes you eligible for specific search features, and it disambiguates your entity so an engine is confident which company it is talking about. Those are real. "Schema gets you cited by ChatGPT" is not established, and the honest move is to say so rather than repeat it.

The retirements confused people here. Google restricted FAQ rich results to government and health sites in August 2023, then retired them completely on 7 May 2026. HowTo went earlier: announced 8 August 2023, gone from desktop by 13 September. Plenty of teams took that as a cue to rip the markup out of their templates.

The display treatment died. The description of the page did not.

• Article schema with honest datePublished and dateModified values, so freshness is legible.
• FAQPage schema whose questions mirror the visible page exactly.
• HowTo schema on genuine procedures, matching the steps a reader can actually see.
• BreadcrumbList, so the page's place in the site is unambiguous.

Your traffic dashboard is now lying to you

AI answers broke the chain from rank to clicks, so a traffic decline no longer means what it used to. A page can be quoted in an answer read by thousands of people and register one visit. Sometimes none.

A team still watching rank-to-clicks will read that as failure, and because the number really did fall, and because somebody has to explain the number in a meeting on Thursday, the page will end up on a list of underperforming content, and then on a list of content to consolidate, and then it will be gone, and the answer engine that was quoting it every day to a few thousand people will quietly start quoting a competitor instead. Nobody will connect the two events. They are four months apart.

• Track citation directly: ask the engines your buyers' questions on a schedule and record which sources get named.
• Expect impressions and clicks to decouple. Stop treating the ratio as a quality signal.
• Weight branded search and direct traffic more heavily, because a citation often produces a name lookup rather than a click.
• Keep publishing where engines actually crawl and quote, which in practice means your own site plus the discussion platforms that get indexed.

Sheevook scores every draft on these levers before you publish, and tells you when the platforms you picked are ephemeral enough that a GEO score there means very little. The scoring is deterministic and runs offline, so it costs nothing and runs on every draft rather than on request.