9 September 2026 · 6 min read
The grouping below is Avid's own, used in every AI Visibility Report we run; the five kinds became five, with comparison and review sites merged into one editorial lane, on 19 August 2026. Both category figures are published research, printed with their house, date and stated market.
A retailer's product page. A four-year-old forum thread. A regulator's guidance note. On the reads we run, the largest single kind of source behind an AI answer is usually one nobody can buy. Between them these pages are writing what a buyer reads about you, and none of them takes a brief.
Which sources are biggest is the question everyone asks, and the answer is rarely one you can act on. The useful question is which sources are both material and reachable, and it answers differently in every category and every market.
90% of the sources AI cites when answering are ones the brand does not control. Foundation Marketing and AirOps, May 2026, reading 57.2 million citations across 5.1 million AI responses covering 50 brands in seven verticals. The market the sample covers is not stated. Read the study.
The reflex is to move the money to other people's pages. Half right. Not controlled and reachable are different properties, and much of what sits inside that figure is not for sale in any form, at any price, to anyone. It is the opening line of a subtraction, not a budget instruction.
We group a read's sources by kind rather than by domain. A list of domains tells you who is winning; a list of kinds tells you what you could do about it. The brand's own site is always its own row, because it sitting low is usually the finding.
Publisher articles with no stake in which brand wins: category explainers, comparisons, reviews, best-of lists, switching and buying guides written by journalists. This is the one lane that takes a brief. A publisher's editorial team can be commissioned to write on a topic, to their own standard, on their own site, which is branded content working as it always has. It is reachable because there is an editor to brief.
Yours. Nothing to buy, which makes it the cheapest lane and the most over-worked one. It carries weight on questions that name you and almost none on the open question that decides a shortlist: a page whose job is to say you are the right choice cannot weigh you against a rival. Fix what is wrong on it, then stop asking it to do the other job.
A product page, a marketplace listing, a retailer's own buying guide. On our reads of retail-sold categories this is regularly the largest kind in the answer, and none of it is a media buy. What reaches it is the trade relationship you already have: range, listing quality, ratings, the accuracy of your product data where the retailer publishes it. The retailer's buying guide is still the retailer's to write, and it will name whoever it thinks is worth naming.
Reddit threads, specialist forums, reviews left by people. It is the layer that reads as the unpaid view, and that is exactly why you cannot place into it. What reaches it is having something specific and true enough that people repeat it, which is a product question wearing a media hat. Read it as a measure of reputation, not as inventory. And treat any claim about its direction carefully: the published evidence on how much of the answer the forums hold disagrees with itself, and we have found nothing that settles it.
Encyclopaedia entries, standards bodies, regulators, government pages, industry associations, academic work. These get quoted for definitions and for anything a buyer treats as fact. None of it is buyable and none of it should be. What reaches it is accuracy: correcting a record that is wrong about you is patient work for someone who knows the institution, not a campaign.
What survives all three is your reachable share, a figure for your category in your market rather than a rule. We have read categories where it is most of the answer, and categories where the editorial citations sit almost entirely with titles no network can place into. In the second case the honest read says so and sends the budget elsewhere: a highlight promising a lane that is not there is worse than none at all.
Slicing editorial into slivers understates it. Comparison, review, switching and guide sites get filed under different labels by any classifier, and separately each looks like rounding. Read as one lane they are frequently the largest reachable kind in the answer. We merged them into a single independent editorial row on 19 August 2026, after a read printed a brand's reachable ground as a sliver when that lane held the majority of its category's citations.
A rival's own site can be filed as editorial. A classifier works on what a domain looks like, and a competitor's brand site or an industry body's consumer site can look like a publication. Check who owns a domain before counting it as reachable.
The highlight lands on the wrong row. Source charts sort by share, and a template carries its accent on a fixed row, so emphasis marks whichever kind falls in that slot. We have shipped a slide highlighting a tiny institutional row while the editorial row beneath it sat grey. Every word was correct; the colour argued the opposite. Check that the row marked as your opportunity is a lane you could place into.
Ask a model about the brand by name and its own site climbs the list. Ask the open question a buyer actually asks, with no brand in it, and the mix tips towards editorial and community. So a count across a whole category is not the count for the questions you care about. Markets diverge as sharply, because publishing industries differ and some large mastheads take a different line on what they will carry.
There is also no file or tag that buys a place in the answer. Google's documentation on its AI features states that "You don't need to create new machine readable files, AI text files, or markup to appear in these features", and sets one condition for a page to be eligible as a supporting link: it has to be indexed and eligible to be shown in Google Search with a snippet. Read Google's guidance. That is a statement about markup, not about money, and it is worth being precise about the difference: nobody sells a slot in the citation set, which is why the one reachable lane is reached by commissioning a publisher.
Three asks of any report, ours or anyone else's. Group the sources by kind, not by domain. Say which kinds can be placed into, and how. Run it on the questions your buyers actually ask, reported separately from the ones carrying your brand name.
The reason to do the arithmetic now rather than next year is that very little is settled. 15.2% of categories have a brand the models clearly recommend. The other 85% are unowned. Semrush with Kevin Indig, in Search Engine Land, July 2026, across 1,094 categories in ChatGPT from January to June 2026, in the United States. Read the analysis.
So the reachable share is smaller than the headline about lost control implies, and more useful, because most of it is unclaimed. Our free AI Visibility Report reads the questions your buyers ask and groups the sources deciding them by kind. Nobody can promise a model will name you. We can show you which of those sources will take a brief.
10 September 2026 · 6 min read
Before any engine can quote your title, the page has to clear conditions that have nothing to do with the writing. The six that decide it, the two halves of a quotable page, and the crawler distinction most robots.txt files have never made.
10 September 2026 · 6 min read
OpenAI's own documentation says a sponsored placement sits below the answer and cannot change it. That makes the ad and the recommendation two different purchases, out of two different budgets.
9 September 2026 · 5 min read
Someone has probably told you to publish an llms.txt file. It is cheap, and it has real uses. It is also a file the largest answer engine says it ignores, so here is the difference between what reads it and what decides your visibility.
See how your brand appears in AI answers, and where trusted publisher content can strengthen its visibility.
