TL;DR: Four systems started reading your website this month and using what they find to decide what you are allowed to do. Meta classifies advertisers into its restricted Health and Wellness tier partly from crawled page language. Google is removing the campaign-level language setting from Search, so eligibility becomes an inference drawn partly from your landing pages. Webflow will generate your agents' instructions from the content already on your pages. And a site-root agents.md is emerging in ecommerce. Meanwhile 68% of US searches end without a click. The test is not a thought experiment: ask ChatGPT, Claude and Google the questions your customers ask and read what comes back, including what it cites. Below is how to find those questions when you do not already know them, why AI Overviews and ChatGPT are two different tests, what the research says about which pages to fix first, and a prompt that runs the whole thing.
Last week we argued that the brand book is the new dev budget: as code commoditises, the money moves to the constraint you hand a model before it builds anything. That post was about what goes in.
This one is about what comes out, and it is the half we underplayed. One line in that piece deserved a post of its own: whatever your site currently says about your brand is what your agents are going to believe about it.
It turns out that is not limited to your agents.
Four things landed in roughly four weeks, and none arrived as an announcement you would have seen.
Meta reads your site to decide what category you are in. Advertisers in health and adjacent fields are being sorted into a restricted Health and Wellness tier based on crawled landing-page content, product descriptions, URL paths and event names, rather than a category anyone selected. The middle restriction tier blocks lower-funnel events, so campaigns keep running while the conversions you optimise toward stop reporting. On sourcing: this is consistent across many first-hand practitioner accounts and documented nowhere by Meta. Field observation, not citable policy.
Google is taking the language setting out of Search. From late September, campaign-level language targeting is removed for Search and AI Max campaigns. Google infers eligibility instead, from ad creative, landing page language and its own read on the user. Another decision moves from a dropdown you controlled to an inference drawn off your pages.
Webflow will write your agent instructions from your own content. Agent Instructions became a per-site context file at Conf 2026, generated from what is already on your pages. Useful, with a sharp edge: an inconsistent site produces an instruction file that encodes the inconsistency and then makes it authoritative.
A site-root agents.md is emerging in ecommerce. Not the repo-root AGENTS.md used by coding agents, which is a settled standard. The website version is newer and far less established: a file telling a shopping agent what you sell, which pages are definitive, to use live pricing, and not to invent claims. Shopify writes one by default. Worth doing, worth not overstating. Same caution on llms.txt, which roughly 844,000 sites have adopted and which no major model provider has confirmed it reads.
Four systems, four reasons, one shared input: the words on your pages.
For twenty years a website was a thing people looked at, and we optimised it accordingly. Hierarchy, hero, scannability, important part above the fold, compress the rest.
That instinct now works against you, because a growing share of the audience does not look at anything. It extracts.
SparkToro put 68% of US searches at zero-click in early 2026, up from 60% two years earlier. Pew, measuring in March 2025, found people clicked a traditional result on 8% of visits when a Google AI summary appeared, against 15% when it did not. The page still did its job in those cases. It did it somewhere you cannot see, for a reader who never arrived.
So the useful model is not a brochure and not a funnel. It is closer to a wiki: the canonical, boring, complete, text-first record of what is true about your company, maintained because other systems read from it and repeat it to people you will never meet.
Here is the case that makes this concrete, and it is the one most founders have not thought about.
Say your products sell through distributors, retailers and stockists across a dozen markets. You have a brand guide. You send assets. You run trainings. And you have always known that messaging degrades as it travels, because a retailer in another country writing a product description has never had your full context and never will.
What changed is who they ask.
That retailer now opens ChatGPT or Gemini to write the social post, draft the product page, or answer a customer question they cannot answer. Is this gluten free? What is the ABV.? Does it ship to Ireland? What is the difference between the two variants? They are not being lazy. They are doing in ten seconds what used to take an email to your export manager and three days.
The model reaches for your website first. That is the good news and the reason any of this matters. If the answer is on your site in text, the retailer gets your answer, and your positioning and product facts travel intact into a market you are not standing in.
If the answer is not there, the model does not stop. It reaches for whatever else it can find: a competitor's comparison page, a three-year-old blog post, a Reddit thread, a distributor's out-of-date listing, a review site that got your specification wrong. That becomes the social post, the product page, the thing told to the customer. Nobody at your company sees it, and the retailer never knows they were handed something you did not write.
This is a first-party data problem wearing a content costume. The information exists. It is in your export manager's head, your spec sheets, your regulatory filings, your sales team's replies. It has simply never been published in plain text on a page a model can read, because nobody thought a page was for that.
Every unanswered question on your site is an invitation for someone else to answer it in your name, to your customer.
The test is not whether you think your page is clear. Go and ask.
Take the questions, run them through ChatGPT, Claude, Perplexity and Google, and read what comes back. Note the answer, note whether it is right, and note what it cites. That takes an afternoon, costs nothing, and replaces an argument about hypotheticals with a list.
Two surfaces, and they are not the same test.
Google AI Overviews sit downstream of Google's index. If you rank for the query you are a candidate for the summary. Classical SEO still feeds this, which is why your existing rankings partly determine what appears. Conductor put AI Overviews on roughly a quarter of searches in its 2026 benchmark.
ChatGPT, Claude and Perplexity run their own retrieval, far less tightly coupled to your Google position. A page that ranks poorly can still be pulled and cited. One analysis after a recent model update recorded around fifteen unique cited domains per response, so citation credit is spreading across more sources rather than concentrating in the top result.
That asymmetry is the whole reason the next section is counterintuitive, and it is why a single "how do I rank" answer does not cover both.
1. FIND the questions, because you probably cannot list them. Almost nobody can name their five hardest customer questions off the top of their head, and the ones people name are usually the ones they enjoy answering. The real list is already written down in six places: the last two hundred emails and chats your support and sales inboxes handled, your Search Console queries, your on-site search logs, your product reviews and returns reasons, the "People also ask" box for your category, and Reddit or forum threads where your brand or product type comes up. For distributed brands add a seventh and treat it as the most valuable: the questions your distributors and retailers ask you. Those are, by definition, the ones their customers are asking them. Cluster all of it and take the top five by frequency, weighted up for anything where a wrong answer costs you a sale or a regulator.
2. ASK the engines and log what they cite. Run those five through ChatGPT, Claude, Perplexity and Google. For each, record the answer, whether it is correct, and the sources. Three outcomes worth separating: the model cited you and got it right, which is the goal; the model cited someone else, which tells you the gap and who is filling it; the model got it wrong, which is the urgent one. Ask it directly what kind of company you are, too. If Meta's classifier is reading your language, a model's answer to "what category is this business in" is a cheap proxy for what a classifier sees.
3. FIX where the answer lives, starting with the pages that are not winning. For every question with no clean text answer on your site, put one there. In a sentence, near the question, in words. Not implied by a chart, not shown in a graphic, not sitting in a video, not behind a contact form. Accordions and FAQ blocks exist so you can do this without wrecking the layout. On sequencing, the Princeton GEO study is the closest thing to real evidence here: adding citations, quotations and statistics lifted a source's visibility inside a generated answer by 30 to 40%, while keyword stuffing scored below doing nothing at all. But the gain is not evenly spread. Adding citations lifted a fifth-ranked page's visibility 115% and reduced the top-ranked page's by 30%. Two caveats that matter: the study tested ranks one through five only, so anything you read about "positions 3 to 8" is someone extrapolating, and the paper is from late 2023 against a GPT-3.5-era engine, so the direction is sound and the decimals are history. Practical version: your winners have the least to gain and something to lose. Start below them.
4. SPECIFY it rather than letting it be inferred. Write the agents.md and llms.txt yourself. State what you sell, which pages are canonical, that pricing and specification must come from the live page, and that claims must not be invented. For distributed brands, publish the things distributors keep asking for as pages rather than as PDFs in a partner portal: full specifications, market availability, certifications, regulatory status by country. A PDF behind a login is invisible to every system in this post. An afternoon's work, unproven and cheap, which is a perfectly reasonable thing to buy.
On cadence: we run weekly site and AEO audits at Forta for exactly this reason. Not because the content is bad, but because the systems reading it keep changing what they reward, and a quarterly cycle cannot track something that moves weekly.
One honest note on effort. NP Digital published 744 articles across 68 sites and found human-written content out-pulling AI-written content by roughly 5.4 to 1 on traffic by month five. Self-reported agency research with no independent replication, so hold it loosely. It points the same way as the Princeton work: specificity and attribution get rewarded, and both need a human to supply them.
Two passes, about an hour total. The first finds the questions, the second tells you what the machines currently say about you. Run the second one yourself in each engine rather than asking one model to guess what the others would say.
You are a market research and AI-search analyst. Your
job is to find the questions my customers actually
ask, then audit how AI engines answer them about me.
Be blunt about gaps. Do not invent data.
My business: [WHAT YOU SELL, WHO BUYS IT]
Markets and channels: [COUNTRIES, DISTRIBUTORS,
RETAILERS, DTC]
My site: [URL, PLUS KEY PAGES]
Regulated angle, if any: [HEALTH, ALCOHOL, FINANCE...]
PASS 1 - FIND THE QUESTIONS
I will paste raw material. For each source I provide,
pull out every distinct question being asked:
[PASTE: recent support and sales emails or chat logs]
[PASTE: Search Console queries, or say SKIP]
[PASTE: on-site search terms, or say SKIP]
[PASTE: product reviews and returns reasons]
[PASTE: questions distributors or retailers asked me]
Then search for: "People also ask" questions for my
category, and Reddit or forum threads where my product
type is discussed.
Cluster everything into themes. Rank by how often it
comes up, weighted up if a wrong answer costs a sale,
a partner, or a regulator. Output the top 10, and mark
the 5 I should treat as critical.
For each of the 10, tell me in one line whether a
correct answer exists in plain text on my site today.
PASS 2 - ASK THE ENGINES
For the 5 critical questions, give me a copy-paste
block I can run in ChatGPT, Claude, Perplexity and
Google, phrased the way a real customer or a retailer
writing a product description would phrase it. Include
one question that asks what category my business is
in.
Give me a table to fill in per engine: the answer, is
it correct, what did it cite, was it me.
Then tell me what to do with each result:
- cited me and correct: leave it
- cited someone else: what page I need and where
- got it wrong: what to publish and how urgently
Output one prioritised action list. Each row: the
page, the exact change, who does it, how long it
takes. Put wrong answers at the top, missing answers
second. No general advice, only named pages.
Ask me one clarifying question at a time when you
need more info.
We run a client-loaded version of this as an onboarding artifact, pre-filled with the brand's proof points, distributor question log and restricted-category language, so the first pass produces a backlog rather than a reading exercise.
The old test of a web page was whether a visitor could find what they needed. The new test is whether a machine, asked a question by someone you will never meet, answers it with your information or with someone else's.
That is a lower bar than it sounds and most sites fail it, because a decade of optimising for scanning taught us to compress, imply and defer. Compression reads as elegance to a person and as absence to a machine. Absence is not neutral. Absence gets filled.
The work is unglamorous. It is writing down the specifics you have been rounding off: the real number, the actual coverage, the plain answer to the question your sales team gets every day and your distributor in another market cannot get at all. It is the least fashionable thing on your roadmap and the cheapest, and it is now load-bearing in four places it was not in June.
Most founder-led companies already have this knowledge. It lives in the heads of the two people who answer customer emails. Getting it onto the page is the entire job.
If you want a second set of eyes on how your site reads to a machine, book a call.
Product and operations are handled. If marketing is the missing piece, let's talk.