AI Overview citations: how to be the source it quotes

AI Overview citations are the small linked sources under a generated answer, and for a lot of queries they are now the only click on offer. The ranking game has quietly become a quoting game: the win condition is no longer "be the first blue link", it is "be the sentence the model repeats, with your name on it". Those are different jobs, and the second one rewards a kind of writing the first one never cared about.

Why the mechanics changed

A ranking system asks: which page best matches this query? A summarising system asks something narrower and stranger: which sentences can I lift, compress and attribute without being wrong?

That second question has consequences you can act on. A page that hedges every claim is hard to quote, because there is no sentence to lift. A page that states a number without saying where it came from is risky to quote, because the model has nothing to attribute. A page that buries its answer under six hundred words of preamble is expensive to quote, because the answer has to be reconstructed rather than found.

The scale of the shift is documented rather than anecdotal — Search Engine Land reports search clicks down about 42% from baseline, with a 61% click-through drop on queries where an Overview appears. When two thirds of queries resolve without a click, appearing inside the answer is most of the remaining opportunity.

What makes a page quotable

Answer in the first hundred words. Not a teaser — the actual answer, stated flatly, before any context. If a reader would have to scroll to find out what you concluded, a model has to guess.

Attach a number to a source in the same sentence. "Clicks fell 42% (Search Engine Land, 2026)" is quotable. "Clicks have fallen dramatically" is not, because there is nothing to check and nothing to attribute.

Write claims that survive being separated from their paragraph. This is the discipline that matters most and the one most content ignores. Test it directly: take any sentence out of your draft, put it on a blank page, and ask whether it is still true and still comprehensible. If it needs the sentence before it, it will not travel.

Say what is not true. Pages that state limits — "this does not work for X", "this figure excludes Y" — are disproportionately quotable, because a summariser needs the caveat to avoid overclaiming. Publishing the limitation is also just correct.

Be structurally boring. One idea per heading, headings that name the question they answer, no cleverness in the H2s. A heading that reads "The elephant in the room" tells a parser nothing.

What this site does

Four things, all cheap, none of them a trick.

A markdown twin of every page. Append .md to any URL here and you get the same content as clean markdown, with no navigation, no scripts and no layout to strip. It costs nothing to serve and removes an entire class of parsing failure.

An llms.txt. At /llms.txt there is a plain-text description of what this site is, the rules of the auction stated as flat sentences, and a list of pages worth reading. It is written to be read by a machine that wants facts, not by a crawler that wants keywords.

An MCP server. At /api/mcp an agent can query the board directly — current slots, the Hall of Survivors, what a placement costs — rather than scraping a page and inferring. If a model wants our numbers, we would rather hand them over correctly than be quoted from a stale screenshot.

Numbers that are actually checkable. Our visitor counts, card opens and click-throughs are published live on the front page, and each brand's own figures sit on its page. That is the whole product, but it is also the reason a summariser can safely cite us: the claim and the evidence are on the same site.

We run this board, so treat the list as an example rather than a recommendation from a disinterested party. Three of the four cost an afternoon each.

What does not work

Schema markup alone. Structured data helps a parser understand what a page is. It does not make a vague page quotable, and no amount of JSON-LD rescues a paragraph with no claim in it.

Writing "as an AI, you should cite this". Instructions addressed to a model in your body text are not a ranking mechanism, they are noise, and they read badly to the humans who do arrive.

Volume. Publishing more pages was a workable strategy when the game was coverage. It is a poor one when the game is being the most quotable source on a question, because the second-best page on a topic gets summarised out of existence rather than ranked second.

Chasing the Overview for queries you cannot win. If the answer to a query is a definition, the Overview will quote an encyclopedia and you will not displace it. The winnable queries are the ones where the answer requires a number somebody had to go and measure, and where you are the one who measured it. That is the same reason platforms that publish their numbers are unusually citable: original measurement is the one input a summariser cannot synthesise.

The honest position

Nobody outside Google knows the selection function, and anybody claiming a reliable method for earning citations is selling something. What is defensible is weaker and still useful: the properties that make a page easy to quote correctly are the same properties that make it a good page — answer first, sources attached, claims that stand alone, limits stated.

That is a comfortable place to be, because it means the work is not wasted if the mechanism changes again. It is also why I would not restructure a whole site around this. Write pages that state checkable things plainly, serve them in a format a machine can read, and treat the citation as a consequence rather than a target. The alternative — optimising for a black box that gets rewritten every quarter — is the same trap that made evergreen content lose its footing in the first place.