AI Reads a Different Internet for Every Industry

What we did

The Arcalea AEO Index scores brand visibility inside AI answers. To do that it collects the answers themselves, across five engines: ChatGPT, Claude, Gemini, Perplexity, and Copilot.

The August 2026 runs covered four unrelated categories: business schools, commercial mechanical contracting, commercial debt collection, and egg donation and surrogacy. Nothing connects those markets. Different buyers, different price points, different sales cycles.

That is what makes the overlap interesting. 9,046 answers, 45,150 citations, 4,384 unique domains.

One disclosure before the findings. About 22% of those citations are bare domain references, where a model named a platform without pointing at a specific page. We counted them, because a model reaching for a domain is still a signal about where it looks, but they are not page-level sources. Every ranking below holds when they are excluded: the same five domains lead, in the same order.

The five that appear everywhere

Counts are host-level, which is the basis the whole overlap analysis uses. Wikipedia here means en.wikipedia.org; adding the other Wikipedia hosts takes it to 460 citations, or 1.0%.

Five domains out of 4,384, and every one of them is a platform where most of the content is contributed rather than published by the organizations being discussed. Reddit is the single most-cited domain in the entire corpus.

That AI engines lean on Reddit is not news; it has been widely reported. What has not been published is the shape around it, which is the rest of this piece.

The qualifier matters. About a third of the LinkedIn citations point at company pages, which are written by the companies themselves. The other two thirds point at posts and articles written by individuals. So the pattern is not purely user-generated, and we are not going to claim it is.

Three further domains technically appeared in all four industries: google.com, ziprecruiter.com, and a financial-markets syndication site. Each was cited fewer than 25 times out of 45,150. We are naming them for completeness and setting them aside as noise.

The obvious question is why a forum outranks every trade publication, every industry association, and every company website when you aggregate across industries. The answer is that AI engines are answering questions, and the questions people ask are the questions other people have already answered in public. Somebody asked which agency to use, and somebody else answered, in a thread, years ago. That thread is the most direct available answer to the exact question being asked.

Almost nothing is universal

The universal five get the headline. The distribution underneath them is the part worth acting on.

98% of every domain AI cited appeared in exactly one industry.

The obvious objection is that this is arithmetic rather than a finding. A domain cited only once can only appear in one industry, and a third of these were cited exactly once. So we tested it at rising thresholds, keeping only domains with real citation volume.

The effect survives. Even among the 70 domains cited a hundred times or more, the ones that actually carry the category, 86% appear in a single industry. This is not a long-tail artifact. AI’s working source set is category-specific most of the way up.

So there is no general-purpose list of sites that make you visible in AI answers. There is a tiny universal spine, and then there is your category, which is almost entirely its own world. Any advice that starts “get listed on these sites” without naming your industry is guessing.

This is also why single-category benchmarks mislead. What AI reads about business schools tells you close to nothing about what it reads about mechanical contractors.

How much the mix changes by industry

Reddit’s own numbers make the point. It appears in all four industries, and its importance varies enormously.

Business schools here combines the M7 and higher-education indexes, the same grouping used for the industry counts above. In the two consumer-facing, high-consideration categories, a forum is the first or second most-cited source in the entire category. In the two B2B industrial categories it is a minor source, and what rises instead is different in kind: the companies’ own websites.

In commercial debt collection, the most-cited domain in the whole category is a single agency’s own site. In commercial mechanical contracting, the top cited domains are LinkedIn and two contractors’ own sites.

That is a real strategic split. In categories where buyers discuss their decisions in public, AI reads the discussion. In categories where they do not, AI reads whatever the companies themselves have published, which means the opportunity to be the source is wide open and almost nobody is taking it.

The clearest example we found

This finding only exists because the index measures AI visibility in two separate layers: what models have internalized without web access, and what they surface when grounded with live search. Collapse those into one number, which is what every other measure of AI visibility does, and the case below is invisible.

One agency in the commercial debt collection index has no AI memory at all. Asked about the category with web access disabled, not one of the three memory-layer engines named it, in any answer, ever. On the memory layer it scores zero.

It ranks first in its entire category on the blended index score.

The reason is visible in the citation data. Its own website is the single most-cited domain in the category, ahead of every trade association, every competitor, and every publication. It has no brand presence in what the models internalized, and it does not need one, because when a model searches, the best available source about commercial debt collection is that company’s own site.

And it is not one page getting cited repeatedly. Those citations spread across 102 distinct URLs on its site. The most-cited of them are not service pages or location pages. They are comparison and best-of pages aimed at specific trades, plus a set of industry pages organised around who the work is for.

That is the pattern in one company. It publishes answers to category questions, so it is the entity available to answer them.

And the control case is in the same index. Another agency in the same category, also with zero AI memory, whose own domain ranks around 100th by citations. Its visibility score is roughly a third of the first agency’s. Same category, same absence from AI memory, opposite outcomes, and the difference is whether their own site is a source worth citing.

You do not need the models to know who you are. You need your website to be the best answer available when they look.

What to do with this

Find your category’s real source list, not a generic one. With 98% of cited domains appearing in a single industry, the list that matters to you is specific to you and it is knowable. Ask five engines your category’s buying questions and note every source they cite.

Take the universal five seriously, and Reddit especially, if you sell into a category where buyers talk in public. Not by astroturfing, which is both wrong and detectable, but by being genuinely present where your category is already being discussed.

If you are in a B2B category where nobody discusses purchases publicly, the opportunity is your own site. That is the harder-sounding option that is actually the easier one, because in those categories the competition to be the cited source has barely started.

Stop treating AI visibility as a single thing to be won. It is a category-specific question with a category-specific answer.

Where to see this for your own category

We publish the underlying indexes instead of describing them. The Arcalea AEO Index carries the live scores, the two-layer split, and the per-signal detail for every entity we track, including the categories in this piece. How the scores are computed is published in full in the AEO Index methodology, and how to read your AI visibility report walks through every section of an index and what to do about it.

If you want the same read on a category we have not published, that is what the index is for. And if you want to start with your own plan instead of your own score, the 21-step Marketing Planning Diagnostic is free and open.

Every figure above is measured, not estimated. That is the point of running the index in the first place.