“So We Start With The Homepage Copy, Right?”
That was the whole discussion. One sentence in a planning meeting, no objections, item closed.
It sounds right, too. The homepage is the front of the building, so of course you clean the front of the building first.
I nodded along and said nothing, which I’ve been doing a lot lately in these meetings.
It bothered me afterwards. Not one person in that room had checked whether the assistants read the homepage at all, and we had just committed a quarter to it on the strength of the fact that it sounded sensible.
Somebody went and counted, as it turns out, in August 2026.
Which Pages Do The AI Assistants Actually Cite?
Deep ones. Out of 3,412 citation URLs collected from five assistants, homepages accounted for 1.5%, while 74.7% sat two or more levels down the path (Foglift, fieldwork 1-10 August 2026, 75 questions and 375 answers).
So what the models reach for isn’t the front door. It’s whichever individual page answers the question one-to-one.
The counting was done by Foglift, a company that sells a tool for tracking and improving citations in AI search. They wrote 75 purchase-intent questions across 25 industries, none of them naming a brand, and put each one to ChatGPT, Claude, Gemini, Google AI Overview and Perplexity.
That produced 375 answers, carrying 3,412 citation URLs across 1,510 distinct domains.
Why The Back Room Beats The Front Door
A model picks its sources to answer the question in front of it, and nothing else.
Given that, a page carrying only the answer to that question is more useful than a page carrying a portrait of the whole company.
Your homepage is a signboard. It says who you are, what you sell, how the last quarter went, and that you’re hiring. All of that is true and none of it is what the person asked.
They wanted one line out of the twenty on that page.
It’s the difference between someone asking for directions and being handed the company brochure. Generous, technically responsive, no help at all.
So the page that gets quoted is the one where that single line is the entire page.
The Domains Differ Wildly From One Engine To The Next
The same study counted something else worth keeping.
Take the top 25 domains for each of the five assistants, put them side by side, and 75% of those domains show up for exactly one engine. The rosters barely overlap.
My first read was that this means five separate programs of work, one per assistant, which is the point where a plan usually dies.
It doesn’t say that. The depth figure above pools citation URLs from all five engines, and the per-engine breakdown of depth isn’t published, so I can’t tell you whether homepages fare better on Perplexity than on Gemini.
What I can tell you is that across every citation the study collected, homepages were 1.5% of them. That holds regardless of which domains each engine happens to favour, and it’s worth settling before anyone budgets for five parallel workstreams.
What To Count On Your Own Site
Start by counting, not writing. How many pages on your site answer exactly one question a prospect would actually ask, and answer it completely?
For most company sites the honest number is low.
The usual shape is a single services page with the answers to eight likely questions buried in it as paragraphs. That page is not, from a model’s point of view, the page that answers any one of those questions.
So the first move isn’t new material. It’s splitting the explanations you already have into one page per question.
Every page you split off lands one level further down the path. That’s the same place the 74.7% is pointing at.
If you also want to know whether you’re being surfaced at all before you worry about which page gets quoted, being found and being chosen are separate problems, and that piece is the better one for deciding what goes where.
How Far This Number Travels
Five caveats. Get them straight before anybody turns 74.7% into a target on a slide.
- Foglift sells a tool for tracking and improving AI citations, so “measure your citations and act on them” is the conclusion that sells the product
- The window is ten days and 375 answers. Model answers move day to day, and nothing here shows the same ratio holding outside that window
- The questions are English-language purchase-intent queries. Citation structure in other languages wasn’t measured
- The industry scores published alongside this (SaaS/B2B 62, e-commerce 48) come from a separate measurement, on its own design and period, and don’t sit next to these figures
- The line about FAQ pages being cited 2.8x more often isn’t Foglift’s measurement — it’s outside research (Aggarwal et al., KDD 2024), not a result of this study
The ten-day window is the one to put on the slide next to any number you take from here.
With all five of those on the table, what survives? The direction does. The pages being quoted were not the front door.
Direction is robust to the ratio moving around a bit, which is more than you can say for the ratio.
Build Question-Level Pages Before You Touch The Homepage
The thing to take away is an ordering.
Polishing the homepage copy can go to the back of the queue. It isn’t where the citations are landing.
What goes to the front is writing down ten questions your prospects say out loud, and giving each one a page of its own. Splitting existing pages gets you most of the way there, and every split adds depth.
Then check whether any of it got quoted — from the model’s side, because your own analytics will never show you which of your pages an assistant cited.
The urge to clean the front of the house first is a real one. Start in the back rooms anyway.
Sources
- Foglift, “AI Visibility Benchmarks 2026”, August 2026 (fieldwork 1-10 August 2026: 75 brand-free purchase-intent questions across 25 industries, put to ChatGPT, Claude, Gemini, Google AI Overview and Perplexity, producing 375 answers containing 3,412 citation URLs across 1,510 distinct domains. Homepages were 1.5% of cited URLs; 74.7% were two or more levels deep in the path. Comparing each engine’s top 25 domains, 75% of those domains appeared for only one engine. Foglift sells a tool for tracking and improving citations in AI search. The industry scores published in the same article derive from a separate measurement whose formula is not disclosed; the “FAQ pages are cited 2.8x more often” figure is cited from Aggarwal et al., KDD 2024, and is not a result of this study.)
