Google says there is no such thing as optimising for AI Overviews. It is half right
Every few months a vendor invents a new acronym for getting quoted by an answer engine, and every few months Google says plainly that no such lever exists. Both statements can be true at once: there is no hidden setting, no markup that requests a citation, and no ranking factor you can buy — and there is still a large difference between a page that gets quoted and a page that does not.
What was actually said
The consistent line from Google is that AI Overviews draw on the same index and the same quality signals as ordinary results, that there is no separate submission route, and that no structured data field asks to be cited. People hear that as "nothing you do matters". What it says is narrower and more useful: there is no shortcut layered on top of search, because the surface is search.
That is a debunking of the products, not of the problem. The acronym industry sells a lever that was never there. The underlying question — when a machine composes an answer in your trade, whose words does it use — is real, measurable, and mostly unanswered by the people who sell the acronym.
Why the work survives the debunking
An answer engine does not quote pages. It quotes passages. Retrieval finds candidate documents, extraction lifts the passages that most directly address the question, and composition assembles an answer from those passages and attributes the ones it leaned on. Every stage after the first is passage-level.
So the thing that changes your outcome is not a tactic bolted onto a page, it is whether any single paragraph on your site is the best available answer to a question somebody actually asks, phrased the way they ask it, and able to stand on its own when lifted away from everything around it. That is ordinary editorial work. It is also exactly what Google has been describing for years when it says to write for people — which is why the company can honestly say there is nothing extra to do and still be describing something most sites have not done.
The test is brutally simple. Cut one paragraph out of your page, paste it into a blank document, and read it. If it needs the paragraph above it to make sense, it cannot be quoted.
How to use the debunking to your advantage
- Competitors who believed the acronym are spending on a lever that does nothing. Competitors who believed the debunking have stopped measuring the surface entirely. The gap between those two mistakes is where the work is.
- Nobody can sell you a guaranteed citation, which means nobody can outbid you for one either. The only currency is being the clearest answer, and a small site can be the clearest answer in its trade far more easily than it can outrank a national brand on a head term.
- Measuring it is cheap and almost nobody does it. Asking the assistants your customers' questions on a schedule and recording which domains come back is a few pounds a month of API calls, and it tells you which eleven pages to get onto rather than how invisible you are.
What we do not claim
An answer from a model's API is not the answer a logged-in consumer sees. Different system prompt, different retrieval, often a different model version. Anyone quoting an API-derived percentage as "your ChatGPT visibility" is over-claiming, and we say so on our own screens: each surface is labelled for what it is — measured through an API, observed in a results page by our own crawler, or not measurable at all. Three honest numbers beat one flattering one.
What to do about it
- Stop buying tactics that claim a lever into AI Overviews. There is not one.
- Rewrite your best commercial pages so one paragraph under a question-shaped heading answers that question and survives being cut out.
- Start measuring which domains the assistants cite for your questions, and treat that list as a list of pages to get onto.
- Refuse any visibility score that does not say how it was measured.