How do you do AEO on a site with thousands of pages?

On a site with thousands of pages you work by selection, not by coverage. Pick the buyer questions worth winning, find or write the one page that should own each, and fix those pages properly. A programme that tries to raise the average quality of ten thousand URLs spends its budget on pages no buyer ever asks a question about.

Selection beats coverage because the citation contest happens page by page, and most of the variation sits between pages, not between sites. Chen and Koudas, University of Toronto, October 2026, in a within-website design over two 90-day collections totalling 158,099 answers and 468,037 citations, found 13% of the variation in citation counts lying between websites and 87% between pages of the same website.

That number cuts both ways on a large site, and both halves are worth saying. It means a strong domain does not carry a weak page, so the big site does not get the win automatically. It also means one well-built page on a large site can be the cited one, which is why a shortlist is a credible plan rather than a compromise.

The practical shape of a first engagement on a large site is therefore an audit that ends in a ranked list of pages and gaps, not a site-wide crawl report. The crawl tells you which templates are broken. The list tells you which twenty pages to change this quarter, and it is the part that needs judgement about your buyers rather than a tool.

Which pages should you fix first on a large site?

Fix the pages that already sit in the retrieval path for a question a buyer asks with money in hand. On a large site there are usually three or four such pages per buying question, and the first job is to decide which one should own it, because two of yours competing is a problem a crawl will not surface.

The reason to start from retrieval rather than from traffic is that retrieval is where most of the decision happens. Chen and Koudas, University of Toronto, October 2026, found that retrieval explained 32% of the within-website variation in ChatGPT's 90-day citation counts, with a 95% confidence interval of 24% to 42%, against 2.8% for all page features combined. Pages the engines already fetch are pages where a wording change can convert a near miss, and pages they never fetch need a different intervention.

Within that set, the fix is wording. The same study found title match and best-paragraph match to the question asked were the only features to pass all five of its robustness tests, at +14.1% (95% CI 10.7 to 17.6) and +15.5% (11.1 to 20.1) citations per standard deviation. One nuance in that paper is worth planning around: for ChatGPT, title match showed no effect at all (-0.2%), which the authors link to ChatGPT rewriting questions into its own searches. The paragraph that answers the question carries more weight than the headline on that engine.

How much a single page can move is the part most enterprise teams underestimate. When we ran our own scorer on one of our own pages, it came back at 28 out of 100, and the rebuild took it to 71 in an afternoon, all of it structure and wording, with no new facts added. The arithmetic of a large site is that twenty of those beat a thousand small template edits.

Does rolling out schema or FAQ blocks across every template move AI citations?

Rolling out schema or FAQ blocks across every template is the most popular enterprise AEO project and the one with the weakest evidence behind it. The measured within-site effect is close to nothing, and in some estimates it points the wrong way.

Chen and Koudas, University of Toronto, October 2026 separated the two comparisons that usually get conflated. Across websites, the features looked good: FAQ blocks +18.1%, summary boxes +20.9%, author boxes +14.9%, schema.org markup +7.5% and page speed +11.7% per standard deviation. Within a single website, none was significant, and the estimates were -3.4% for FAQ blocks, +6.9% for summary boxes, -4.6% for author boxes, -5.3% for schema.org markup and -1.1% for page speed. At equal retrieval rank, author boxes, FAQ blocks and schema.org markup went with slightly lower odds of being cited.

The explanation the authors give is Simpson's paradox: those features cluster on the kind of site that gets cited more for other reasons, so a cross-site correlation picks up the site instead of the feature. They present their results as associations, not causal effects, and one preprint is not a literature, so the sensible reading is not that markup is harmful.

What the finding does change is what you expect from the rollout and what you pay for it. Markup is cheap, it costs nothing to keep, it carries real benefits elsewhere, and it belongs on the templates. It is not a visibility programme, and a supplier who proposes a site-wide schema deployment as the quarter's AI visibility work is selling the cheapest thing to automate instead of the thing that moves the number. Our page on whether schema markup helps AI citations goes through the narrower case where it genuinely matters.

How many pages does an AI visibility programme actually need to touch?

Count buying questions, not pages. In most B2B categories the questions that produce a vendor answer number in the dozens, so a first programme touches dozens of pages, whether the site has four hundred or forty thousand. The page count is a fact about your CMS rather than about the work.

Two measurements keep that scope honest. Chen and Koudas, University of Toronto, October 2026, found that about 75% of the variation in citation counts within a website was repeatable but explained by none of the page features they tested, which means a page score tells you whether a page is in a condition to be used, not how often it will be cited. And their drift measurement found that the repeatable part does not hold still: between two sampling windows ten weeks apart, only 14% of it was shared on ChatGPT, 29% on Gemini and none on Perplexity, so a programme that spends two quarters rewriting a template library has been working against a target that moved repeatedly while it ran.

There is a second reason to keep the list short on a large site. Two pages of yours answering the same question split your own retrieval chances and neither becomes the obvious one to cite, which is a condition large sites produce by accident through years of campaign landing pages. Consolidation is often the highest-value early move, and it reduces the page count instead of adding to it, as two pages competing for the same AI answer covers.

So the honest scope conversation for a large site runs: here are the 40 questions, here is the page that should own each, here are the 12 that need writing and the 28 that need rewording, and here is what we are deliberately not touching. A proposal that instead scales with your URL count is priced off your CMS. What drives the real number is in what makes one AEO quote higher than another.

What does a prioritisation list for a large site look like?

A prioritisation list on a large site sorts pages into five classes, each with a different fix and a different cost. The classes matter more than the order, because they tell you which team does the work and whether the page needs writing, rewording, merging or leaving alone.

Page class	How you spot it	Why it is on the list	The fix
Retrieved and passed over	Appears as a source on some runs of a buyer question, never in the answer	The expensive half, being fetched, is already done	Reword the title and the answering paragraph to match the question
Question with no page	A buying question where nothing of yours is retrieved at all	An absence rather than a weakness	Write one page, scoped to that question
Two pages, one question	Several URLs of yours answer the same thing, none completely	Your own pages are splitting the chance	Merge into one, redirect the weaker URLs
Blocked or unrenderable	Fetching the page as a crawler returns a shell or a wall	No content question can be answered until this is clear	A technical fix, usually one sprint, upstream of everything else
Long tail with no buyer	Thousands of pages nobody asks a buying question about	It is where a coverage-driven programme spends its budget	Leave it alone, keep the templates sound

One line on that table is the one that gets argued about, and it is the last. Leaving thousands of pages untouched feels like an incomplete job, and on a large site it is the decision that makes the rest of the work affordable. A page that answers no question a buyer asks cannot be cited for one, however well it is marked up. What to settle before any of this starts is in what to fix before hiring an AEO agency.

How do you measure progress on a large site without fooling yourself?

Measure a fixed prompt set against a baseline you took twice, and read the page-level reports as confirmation, not as the result. On a large site the temptation is to track a site-wide average, which moves for reasons that have nothing to do with the work and hides the twenty pages you actually changed.

Google's own reporting now covers its AI surfaces, with a stated limit. The generative AI performance report includes impressions for AI Overviews and AI Mode, can be grouped by pages, countries, dates and devices, and excludes data from Search Labs experiments, per Google's documentation for the generative AI performance report, read 22 October 2026. Impressions tell you that your pages were shown on those surfaces. They do not tell you which answers named you or which pages were quoted.

That is why the prompt set stays the primary instrument. Run the buying questions on a schedule, record every cited URL with a date, and compare runs with each other instead of reading one. We baseline the same set twice, a week apart, before anything changes, so the first comparison contains your own noise instead of mistaking it for movement. The arithmetic, including how to size that noise, is in how to measure AI citation decay.

Report it as a range and per page class. Twenty pages moving from retrieved-and-skipped to cited is a readable result; a two-point rise in an average across eleven thousand URLs is not, and on a large site the second is what gets presented when nobody agreed the scope. If you want the shortlist for your own site before committing a budget, our free visibility assessment produces it from your buyers' questions rather than from a crawl.

What else do people ask about AEO on large sites?

The five questions below come up with teams running large sites, where the page count itself is the thing that makes the project look impossible.