Map Your Citation Surface: Every Page Where AI Can Find You

Ask ChatGPT to recommend a project management tool and watch where the answer comes from. It rarely quotes the vendor's homepage. It quotes a roundup on a software review site, a Reddit thread where three people argue about Notion, a comparison page some affiliate blogger wrote in 2022. The brands that get named are the brands that already appear, described favorably and in retrievable language, on those external pages. Your own marketing site is one voice in that chorus, and usually not the loudest.

This is uncomfortable if you have spent a decade optimizing your own domain. The lever you control best is the one the model reaches for least. Observed citation mixes often lean heavily toward third-party pages, but no universal percentage has been established. Measure the split for your own category and engine set. The pages that decide whether an assistant recommends you are mostly pages you do not own.

So the optimization target moves. Instead of a site you tune, you have a surface you map. We call it the Citation Surface: the full set of external pages where you co-occur with the queries your category attracts, the listicles and review sites and forum threads and docs and Q&A answers that an assistant might pull into its context window when someone asks about what you sell. This article is a build-along. By the end you will have run a five-step pass over your own category, and you will know which of those external pages are working for you, which are wrong about you, and which do not mention you at all.

What the Citation Surface actually is

Start with the mechanism, because the surface only makes sense once you know what the model does with a prompt. A retrieval-augmented system takes a question, rewrites it into several sub-queries, and sends each one to a search backend. ChatGPT uses provider-managed web retrieval, Perplexity operates its own crawler and a provider-managed retrieval stack, and some systems query a vector store. Each sub-query returns a candidate set of pages. A reranker trims the pool to a few passages, those passages go into the context window, and the model writes its answer from what it sees there.

Every page that can enter one of those candidate sets for a query in your category is part of your Citation Surface. Not every page on the web, and not only your own pages. The specific external URLs that already rank, already get crawled, and already contain language close to the sub-queries your buyers trigger. Your homepage might be one of them. The G2 category page almost certainly is. That old Reddit thread might be too, depending on how the engine weights forum content that week.

The surface has a useful property: it is finite and it is enumerable. There is a bounded set of pages that realistically show up for "best CRM for small teams" and its cousins, and you can go find them. That is what makes the work tractable. You are not trying to influence the whole internet. You are trying to influence a countable list of pages that a specific set of prompts keeps surfacing.

Step one: enumerate your prompt-space

You cannot map the surface until you know which prompts it answers to. This is the same prompt-space work that replaces flat keyword lists, and if you have already built a Prompt-Space Coverage map you can reuse it directly. If you have not, do a fast version now.

Write down the real questions a buyer brings to an assistant when they are close to choosing something in your category. Pull them from three places: recorded sales calls, support tickets, and your own team typing honest questions into ChatGPT and Perplexity. Do not compress them into keywords. "What's a good Mailchimp alternative that doesn't get expensive at 50k contacts" is one prompt, and it will fan out very differently from "best email marketing tools." Group what you gather into a few dozen intent clusters: category-shaped ("best X for Y"), comparison-shaped ("X versus Z"), and problem-shaped ("we keep hitting this, what fixes it"). Aim for thirty to fifty representative prompts across those clusters. That is enough to expose the surface without drowning you.

Step two: capture the URLs each engine cites

Now run those prompts and record where the answers come from. For each prompt, submit it to the engines that matter to you, the four being ChatGPT, Claude, Gemini, and Perplexity. Perplexity and ChatGPT's search mode expose their sources most readily, so start with those two before moving to the engines that keep their retrieval hidden. Copy every cited URL into a spreadsheet alongside its prompt, engine, and date.

Two details make this capture worth doing rather than doing sloppily. Run each prompt more than once, because retrieval is not deterministic and a single run tells you what happened that time, not what tends to happen. And log the source URLs even for prompts where you do not appear, because a page that gets cited for your category while never mentioning you is exactly the page you most need to know about. After thirty prompts across four engines with two runs each you will have a few hundred rows, many of them pointing at the same two dozen domains. That repetition is the signal.

Step three: cluster and rank by co-citation frequency

Collapse the spreadsheet. Group the captured URLs by page, then by domain, and count how often each one appears across your whole prompt set. The pages that show up again and again, across different prompts and different engines, are the load-bearing ones. A review-site category page that gets cited for eleven of your thirty prompts matters more than a blog post that surfaced once for one obscure phrasing. Rank the whole list by that co-citation frequency.

This ranking is the core of the method, and it is worth slowing down on why frequency is the right sort key. A page cited across many of your prompts sits in many candidate sets, which means the reranker keeps finding it relevant to your category as a whole rather than to one narrow question. Influencing that page changes what the model sees for a large slice of your prompt-space at once. A page that surfaced for a single prompt might be worth a look, but it moves one cell of your coverage map. Sort by reach, not by how much you happen to dislike what a given page says.

Frequency of citation is a downstream proxy for something upstream, and it helps to name it. What you are approximating is Share of Retrieval, the fraction of candidate sets your content lands in before anything gets cited. You cannot see the candidate sets directly from the outside, but a page that gets cited constantly is a page that enters candidate sets constantly. Ranking by co-citation frequency is the best view of the retrieval layer you can assemble without instrumented access to the engines.

Step four: audit how you are represented on each page

Take the top slice of that ranked list, maybe the top twenty or thirty pages, and open every one. For each, answer a plain question: how does this page represent you? There are four states worth tagging, and each points to a different action.

Absent. The page ranks for your category, gets cited for your prompts, and does not mention you at all. This is the most common finding and often the most valuable, because these are the rooms where buyers are being advised and you are not in the conversation.

Present and accurate. You appear, the description is fair, the positioning is roughly what you would write yourself. Leave it alone and note it as an asset. These are the pages carrying your visibility right now.

Present but outdated. You appear, but the pricing is two versions old, the feature list predates your last three releases, or the page still calls you a "startup" three years after Series B. Cited sources skew old in general, with one analysis putting the average age of a cited page around seventeen years, so stale-but-ranking is a normal and fixable state.

Present but inaccurate. You appear, and something is wrong in a way that hurts: a limitation that no longer exists, a comparison that gets your core feature backward, a security claim that was never true. These are the highest-priority fixes because the model will faithfully repeat the error to a buyer.

Add one column while you are at it: can you influence this page directly, or must you earn the placement? A page you can edit or submit to sits on one side. A page you can only influence by being genuinely worth writing about sits on the other. That distinction drives everything you do next.

Step five: act, and keep the two levers separate

The pages split cleanly into two piles, and conflating them is the most common way this work goes wrong.

The direct pile is where you have a mechanism to change the page. Your own documentation and comparison pages, obviously. Review platforms like G2 or Capterra where you claim your profile and keep it current. Directory and listicle sites that accept vendor submissions or corrections. Wikipedia and similar references where a factual correction with a source is welcome and a promotional edit is not. Forum threads and Q&A posts where you can add a genuinely useful answer under your own name. On these, the work is mechanical: claim the profile, submit the correction, update the stale figure, answer the real question. Fix the inaccurate and outdated pages first, because a wrong page cited often is doing active damage, then work down the absent pages you can get listed on.

The earned pile is where you cannot edit the page and should not try. An independent reviewer's roundup, a journalist's comparison piece, the deep-dive some unaffiliated blogger wrote on their own time. You get onto these the slow way, by being worth including: shipping the thing that makes you belong in the roundup, giving reviewers real access, doing work that earns the mention. This is slower and it is the more durable half, because a page that chose to include you on merit is harder for a competitor to dislodge than a directory row you submitted.

There is a line you do not cross, and the Princeton GEO research draws it for you. When researchers tested what actually raises visibility in generative answers, the methods that worked were adding citations, adding statistics, adding direct quotations, and improving fluency, up to roughly a 40 percent lift on their visibility metric. Keyword stuffing did nothing or slightly hurt. Every method that worked amounts to making a page more genuinely informative and better sourced. Getting corrected where you are wrong and getting included where you genuinely belong is the whole game. Planting fake reviews, spinning up sockpuppet forum accounts, or paying for placements dressed as independent opinion is a different activity, and beyond the ethics of it, the engines and the platforms are getting better at discounting exactly that kind of manufactured signal. Accuracy is the strategy. Manipulation is a liability with a delay on it.

Where this sits in the funnel

Map the Citation Surface once and you have a snapshot. The reason it needs to become a habit is that the surface moves. New review pages get published, forum threads age out, engines change how they weight sources, competitors get themselves added to the roundups you are absent from. The gap between the pages that could answer a query in your category and the smaller set actually retrieved, the Retrieval Gap, is not a fixed distance. It widens every time someone else earns a placement you did not.

The surface work targets the middle of the AI Visibility Funnel, the four stages running Retrievability, Candidacy, Citation, and Attribution. Enumerating prompts and capturing citations measures Candidacy and Citation directly. Fixing the pages improves Retrievability upstream of both. What no manual pass gives you is continuity, because a spreadsheet you built in March describes a surface that no longer exists by June. Running thirty prompts across four engines twice a month by hand is the kind of task that gets skipped the first busy week. Tools like shareof.ai↗ exist to keep the capture and the ranking current so the map stays alive between your audits, but the framework is yours to run with a spreadsheet and an afternoon, and running it once by hand teaches you more about your category than any dashboard will.

A note on measurement

Two honest caveats before you start. The third-party share figure, the seventeen-year source age, the correlation between branded search and citation: these are practitioner estimates from specific datasets, and your category may sit well off the average. Measure your own surface rather than trusting the round numbers, because the whole point of the exercise is that the answer is different for every category. And retrieval is noisy, so a page cited today may be quiet next week. Weight your priorities toward pages that show up consistently across runs and engines, not the one that surfaced in a single lucky query. The list you build this week is a hypothesis about where AI finds you. The pages you fix, and the ones you finally earn your way onto, are how you test it.

Primary sources and evidence

• Web search in the OpenAI API↗ — current search-tool behavior and citations.
• OpenAI web crawlers↗ — official user-agent controls for search and training.

Continue through the system

• Reverse-Engineering Co-Citation: Find the Pages That Decide Your Category
• What to Do When a Competitor Owns the AI Answer for Your Category
• The AI Visibility Funnel: A Diagnostic for Where You're Losing Citations

Run the practical layer with the free AI visibility scan, explore the AI search library, or see the AI visibility benchmarks.