AI Citation Strategy: How to Get Into the Pages Answer Engines Read

Most GEO advice stops at "get mentioned on third-party sites", which is true in the way that "score more goals" is true. The useful version of the question is which sites, in what order, and how do you know. This guide is the answer I wish I'd had eight months ago, and it is mostly about counting things rather than writing things.

The six posts linked at the bottom go deep on individual steps. This one is the map.

What a citation strategy actually is

A grounded assistant does not hold opinions about brands. When someone asks it for the best tools in your category, it runs a search, pulls back somewhere between three and fifteen pages, reads them, and writes a summary of what those pages say between them. The names in the answer are the names in the pages.

So the strategy is a list of URLs. Not themes, not pillars, not a content calendar. A list of specific pages that keep supplying your category's answers, sorted by how often they get used, annotated with who owns each one and whether you are already on it. If you cannot produce that list, you have opinions rather than a strategy, and you are about to spend a quarter guessing.

I say this with some feeling because I spent most of last spring writing blog posts into the void on the theory that good content wins. Our own site sits at average position 68 in Search Console for unbranded queries. Nothing about writing more was going to change that.

Build the frequency table first

Write down twenty to thirty questions a buyer in your category would actually type. Full sentences with constraints in them, because that is how people talk to an assistant. "Best analytics tool" is a search query left over from Google. "What analytics should I use if I'm on Shopify and I can't write SQL" is what gets typed into ChatGPT, and it is the second kind that has an unclaimed answer sitting behind it.

Run every question through a grounded engine and record, for each answer, which URLs were cited and which host each one sat on. Perplexity is the easiest place to start because it cites openly on every query. Then run the whole set again two days later. The retrieval moves between runs, sometimes a lot, and one pass will tell you something confident and wrong.

Now count. Group the URLs by host, sum how often each host appeared, and sort descending. That table is the strategy document. Everything below is how to read it.

One trap worth naming, because it cost me six weeks. Record whether each answer was grounded, meaning the engine genuinely searched rather than answering from memory. Ungrounded answers name brands with no sources behind them and look identical in the chat window. I once had 236 stored URLs I believed were citations. 213 of them came from an ungrounded model and were just domains it had memorised.

What the table tends to look like

When we did this for our own category we ended up with 533 cited URLs across 360 unique domains. Two things in that dataset surprised me enough to change what we do.

The first is format concentration. Sorted by what kind of page each URL was:

Three formats are 73.6% of everything. Everything after that is a rounding error, and some of the rounding errors are things the industry talks about constantly. Reddit is real at 6.5% and it is about a tenth of the listicle surface. I run a company that started life as a Reddit tool, so writing that sentence took a while.

The second surprise is ownership. About 44% of the corpus belonged to editorial sites that nobody in the category controls, roughly 22% to competitors, and the rest split between corporate, institutional and social. The editorial 44% is the pitchable surface and it is the largest single block. The competitor 22% cannot be pitched and has to be displaced by out-publishing it.

And the concentration is steeper than a long tail suggests. Our top 30 URLs carried about a sixth of the total citation weight across all 533. The single most-used page in the whole category accounted for roughly 9% on its own. Twenty placements is a realistic quarter of work and it beats a year of publishing on your own domain, which is an uncomfortable thing to conclude after a year of publishing on your own domain.

Sort the list into four piles

Once you have the table, every URL on it goes into one of four buckets, and the bucket decides the action.

Pages owned by genuine publishers with an editorial process are the primary target. Large platforms with open contributor programmes belong here too, even when they look like competitors. A pitch to a rival's "10 best tools" listicle is a legitimate ask surprisingly often, because the person maintaining that page is judged on the page being useful.

Pages owned by individuals are the fastest of the four and the most overlooked. In our data, two of the six most-cited URLs in the entire category were single articles published by individuals on LinkedIn. No editor, no queue, no gatekeeper. Anyone reading this could publish a comparable page this afternoon.

Pages owned by direct competitors are displacement work. You are not getting added, so the question is whether you can publish something better on the same question and take the retrieval slot. Our own comparison pages do get cited, which tells me the mechanism works in both directions.

Forums and communities are the fourth pile, and they are worth less than their reputation. Real, but small, and the tone rules are strict enough that a clumsy attempt costs more than the citation is worth. I have the bans to prove it.

The order to work in

Qualify before you pitch. A page that gets cited twice in thirty questions is not worth an email, and the frequency table is what tells you which is which. How to get added to a best-tools listicle covers the qualification test and the four kinds of list owner in detail.

Start with the pages that have no gatekeeper, because they convert in hours rather than weeks and they teach you what a citable page looks like before you spend your one shot on a publisher.

Pitch the editorial pages next, hardest and highest-frequency first, while you still have patience. The email itself matters more than people expect and the templates I use are here, including the lines that get you ignored.

Then re-run the frequency table monthly. Placements take weeks to show up in retrieval, the corpus itself shifts underneath you, and without the re-run you cannot tell the difference between a placement that worked and a good week.

What this replaces

If you have been treating AI visibility as an on-page problem, the frequency table will probably reallocate most of your budget. Schema markup, llms.txt, crawler permissions and clean HTML all matter, and all of them are about not being excluded once an engine reaches you. None of them give an engine a reason to reach you.

We ran that experiment on ourselves in public. Four mechanical fixes to one page moved its AEO audit score from 58 to 75 within minutes of deploying. Whether that page gets cited any more often than it did before is a separate question with a separate answer, and anyone selling you the score as a proxy for the citation is selling the easy half.

The frequency table is the hard half. It is also the only part I have seen actually move the number.