What are AI assistants actually reading about your brand?

How AI Answers Questions About Brands · Part Two of Four

Part Two: The Source Study

2,357 cited pages audited · 21,213 citations traced · 10 national brands

First published September 2, 2026 · updated September 3 with a fresh full run — the study now measures five engines, adding Google’s AI Overview.

The series
Part One · The Retrieval Study — do the engines even search before answering?
Part Two · The Source Study — you’re reading it
Part Three · The Reddit Pipeline — the most negative layer of the study, and the two engines that carry it to buyers
Part Four · The Agency Playbook — what to fix first, and how much holds still from one day to the next

When a buyer asks an AI assistant about a brand — is it worth it, which one’s best, is it legit — the assistant usually goes out to the web, picks a handful of pages, reads them, and writes its answer from what those pages say. The pages show up under the answer as citations. Most people never click one. Almost nobody asks the bigger question: what’s actually on them?

It matters because an AI answer isn’t a ranking. A Google ranking is a position on a list you can see — type the search, count down the page, there you are. An AI answer is composed, written from a stack of pages the engine chose and read on the buyer’s behalf. Your answer is only as good as the engine’s reading list, and nobody gets to see the reading list. Part One of this study measured how often the engines search at all; this part is about what they read when they do.

So I pulled the reading list. Across ten national brands — Gymshark, YETI, Hoka, Helix, Peak Design, AG1, Sonos, Patagonia, OLIPOP and Caraway — the five engines — ChatGPT, Claude, Gemini, Perplexity, and now Google’s AI Overview — put 21,213 citations behind 2,500 live answers, pointing at 8,247 different pages. From those, the audit selected the fifty or so most load-bearing pages per question type for every brand: 2,357 pages. I sent the instrument after every one of them.

“Read” is a strong claim, so here is exactly what it means. Every page the instrument could open went through its reading layer: the full text, the entire comment tree when it was a forum thread, the complete transcript when it was a video — with up to thirty-one judgments recorded per page. Where does the brand appear, and how prominently? What does the page actually say about them? Which rivals share it? Who wrote it, and when? Did it even open? The machine read every word. I read the record it produced.

Here’s what a single entry in that record looks like — AG1’s cited YouTube videos, exactly as the reading layer logged them:

Finding 1 · One website tops every brand’s reading list: Reddit

Start with the simplest question you can ask about a reading list: which websites keep showing up? Ten brands, ten different industries — cookware to coolers to mattresses — so you’d expect ten different rosters. Mostly, you get one.

Five websites made every brand’s list this wave. But the top of it isn’t close: Reddit, cited three times more than any other site on the web. Here are the sites that recur across the study, counted from all 21,213 citations:

• reddit.com1,578citations — all ten brands, triple the next site
• youtube.com533all ten brands
• amazon.com244all ten brands
• forbes.com221nine of ten brands
• walmart.com191all ten brands
• trustpilot.com169all ten brands

Look at what that list actually is. A forum. A video site. Two stores. A review site. The places your buyers talk to each other and the places that rank things. What’s not on it is any brand’s own website — the surface most marketing budgets are spent polishing. The engines’ reading list is overwhelmingly about you, and overwhelmingly not by you.

The roster hides a second story: the five engines are not eating from the same table. Take Reddit, the site that tops every list. For Perplexity, Reddit is a seventh of its entire diet — 1,208 of its citations. For Google’s AI Overview it’s nearly an eighth — the second-biggest Reddit appetite in the study. For Gemini it’s about four percent. For ChatGPT, two citations in a thousand. And Claude? Claude cited Reddit exactly zero times — in 7,260 citations. Same questions, same day, same ten brands.

The engines don’t even agree on how much to show you. Here’s the median number of sources each engine displayed with an answer:

• Perplexity16half what it disclosed in the first wave — the engines change how much they show, too
• Claude16
• Gemini5
• Google AI Overview5references per shown Overview
• ChatGPT0the median ChatGPT answer shows no sources at all

For anyone doing this work professionally, this is the finding to sit with. There is no single “AI diet” to feed. The page your brand lives on for one engine can be invisible to another, and a source strategy built by looking at only one engine’s citations is a strategy for a fifth of the market. Here’s Reddit’s share of each engine’s diet, drawn out:

Reading 1,884 pages tells you something counting them never could: what kind of pages the engines actually lean on. The reading layer classified every one, and the shape is lopsided. The biggest single slice of everything the engines read is a ranked list — “best cookware brands,” “top 10 coolers,” somebody’s round-up with your brand at position four. Another fifth is forum threads: your buyers talking to each other. And a full fifth is now video — up sharply with Google’s AI Overview in the mix, which cites YouTube more than any other site. Traditional editorial — an article by a journalist, about you — is fifteen percent.

One brand breaks the pattern, and it’s worth naming: OLIPOP is the only brand in the study whose reading list has more editorial than ranked lists. That mix is earnable. Nine brands out of ten just haven’t earned it.

The full composition of the 1,884 pages the instrument read:

Every page the instrument read, classified by what it is. Half the pages are a ranked list or a forum thread — and another fifth are video, the layer Google’s Overview leans on hardest.

Here’s where reading beats counting. A citation count says a page “mentions” your brand. But “mentioned” can mean the whole page is about you — or one line in a list of six. The engines cite both kinds, and only the reading layer can tell you which one you got. On every page where it could be measured — 712 of the read pages — the instrument recorded exactly how the brand appears. Three real entries from the record:

Part One showed the engines writing years nobody asked for into their own searches. This is the other clock — not what they searched for, but how old the pages they cite turned out to be. Of the pages that carry a readable date — 1,330 of them; another 967 carry no date at all and 60 more carry one that doesn’t parse, and that’s said plainly in the methodology — more than half are over a year old, and a third are over two.

And the age isn’t spread evenly. It lives in the forums. Nearly half the forum threads the engines cite are more than two years old; for ranked lists it’s one in six, because list publishers refresh for a living. The conversation layer of your reading list — the part where buyers talk to each other — is the part most likely to be describing a product you’ve since fixed, repriced, or discontinued.

Two entries from the record. Neither page is wrong to exist — the finding is that engines still hand them to buyers as if they were news.

The last finding is the one I didn’t see coming. To read 2,357 pages you first have to open them, and 353 — 15.8% of the pages that returned a verdict — wouldn’t open at all for our machine reader: blocked, dead, or gone. (As captured the count was higher; a batch of failures carried the signature of a capture fault and is excluded rather than blamed on the web — the methodology below discloses both numbers.)

But the group that failed hardest is the one nobody would guess. The brands’ own websites: 77 of the 143 cited own-site pages — 53.8% — wouldn’t open, 70 of them because the site’s bot protection turned our machine reader away. Dedicated review platforms were worse: 14 cited pages, 14 blocked, a perfect score. The pages brands control most tightly are the pages a machine reader is least able to read.

One page makes the point better than any rate. One of the ten most-cited pages in the entire study is a Forbes greens-powder round-up — 35 citations this wave, and every one of the five engines leans on it for AG1 answers, Google’s AI Overview included. Open it in a normal browser and it loads fine — it names AG1 as its top pick. But send any machine reader, ours included, and Forbes turns it away at the door. Humans can read it. Machines mostly can’t. The engines cite it anyway. Whatever route they take to it, an automated audit can’t follow them in.

The obvious next question — the one I asked myself — is whether the engines could read those pages either. That can’t be claimed, and I won’t. Engines fetch wearing different badges: a site that blocks one crawler may wave another through, some engines hold privileged access to publishers who block everyone else, and some of these pages were alive when the engine first read them and died later — the first wave’s deleted Reddit threads proved that path. What can be said without qualification is narrower, and it still matters: a citation behind a machine wall can’t be verified by any automated reader — not an audit tool, not a monitoring service, not an AI assistant asked to double-check its own source. A person can still click through; everything built on clicking by hand doesn’t scale past a person. And in the deleted-thread cases, there’s nothing left for anyone.

And two brands prove it’s a choice, not physics. OLIPOP’s own site let our reader through on all 15 of its cited pages — a perfect record. AG1’s blocked all 22 of its own. Same study, same reader, same day. Whether a machine reader gets into your site comes down to configuration — the bot protection in front of it and the robots.txt file behind it. Both are settings, both are checkable in minutes, and for a lot of brands both were set by a security vendor’s default rather than by anyone deciding how the site should treat an AI reader.

“Wouldn’t open” = our machine reader couldn’t fetch the page at read time. A person in a browser can often still get in — the wall is for automated readers. Not a claim about the engines’ own access.

Fetch your own homepage the way an AI crawler does — curl -I -A "PerplexityBot" yourdomain.com — and look at the status code. A surprising share of brands will watch their own front door answer with an error.

The full version is below: all ten brand reports, every cited page, what it is, what it says, and whether it opened. This is a dated snapshot of a moving system — if you run it on your own clients and the shape comes back different, tell me. How much the system moves is Part Four’s question.

Forty years in the creative industry teaches you where the money goes: onto the surfaces a brand controls. The website, the campaign, the photography. What this part of the study says, as plainly as I can put it, is that the surfaces deciding your AI answers are mostly surfaces you don’t control and have never read — a reading list that is one-quarter ranked lists, half over a year old, and partly unreadable even to the people it’s about.

Part Three follows the roster to its strangest entry: Reddit — the most negative layer of the whole study, and a pipeline that had one carrier in the first wave and has two in this one.

Glenn · Founder, styleforge.io · September 2026

The complete audit for every brand in this article is published below — the same ten reports from Part One, generated by StyleForge’s AI Visibility Pulse, and each one runs 200 to 250 pages. The reading layer this article is built on fills its own chapter in every report: every cited page, what kind of page it is, how old it is, what it actually says about the brand, and whether it opened.

Click any brand to open its full report. If you run an agency, read one the way your client would — this is the deliverable, the thing you hand across a desk.

Each of these reports is white-labeled — generated under the agency’s own name and presentable straight to the client, and you control which sections every report includes. The agency name on these ten is a stand-in.

Every check in the study, written out in full.

Related resource: the complete guide to earned media — definition, examples, and a working earned media value calculator. The coverage this series measures is earned media; that guide is how to value it.