Second order thinking
AI is an audience, not a channel
Judge AI by the visitors it sends, and you will conclude it does not matter. Judge it by what it reads, and you will find an audience you have never mapped.
Written by Mike Barrett & Sam Bailey
13 Oct 2026
Written by Mike Barrett & Sam Bailey 20 August 2026
In September, AI assistants fetched 26,500 pages from the eight client websites we monitor. Each fetch happened in real time, because a person had just asked a question. For comparison, humans viewed 120,000 pages on the same sites over the same month. That is roughly one live read for every five pages a person viewed.
That is the portfolio average, but it doesn't tell the full story. On the one high-traffic, broad-audience site in the set, live AI reading is still a small fraction of human viewing: one read for every twenty page views. On the specialist sites, the ones whose whole marketing is built on thought leadership, it runs at between a fifth and three-quarters of human viewing.
Live reads: Why they are important and how we measure them
On our own site, the live assistants fetched more than one page for every three a person opened. If you include the training and index crawlers, the machines read more than six pages for every one a person did.
It is still a relatively small sample at eight sites (though we are adding to it), but so far the patterns we cover in this article are consistent across each of the sites. So we trust them, although it is still too soon to be able to call out any specific trends. The numbers are also narrow on purpose. We count only live reads: the fetch that happens because a person asked an assistant something and it reached for your page.
Our previous article included the operators' search-index crawls as well; they belong on each operator's journey, but they rise and fall for reasons we cannot explain from the outside, so they are not in these figures. And we count only reads whose identity we can verify against the operators' published servers; in September, about one per cent of the requests claiming to be these assistants failed that check, and they are left out.
We also leave out client sites behind Cloudflare, where the cache serves responses to machines before the origin server sees them, so the logs undercount; we are actively working on bringing that data in. Read everything that follows as our observations of the patterns and the baseline, rather than a benchmark.
Live AI reads per 100 human page views
Pages fetched live by AI assistants because a person asked a question, set against GA4 page views, on eight client websites.
September 2026
Specialist, thought-leadership sites
High-traffic, broad-audience site
On the sites built on thought leadership, assistants fetch a page live for every one to five a person views. On the high-traffic, broad-audience site it is one in twenty.
Eight client websites not behind an edge cache, 1 to 30 September 2026. Live fetches only (no index or training crawls), verified and plausible bot identity only; about one in a hundred requests claiming to be these assistants failed verification and is excluded. Sites anonymised.
Most marketing teams have no idea which side of the line they are on, because the data in their reporting is not set up to answer the question. AI shows up in their world as a traffic source, a row in the channel report that is typically less than 1% of visits. If you judge it using that data, it looks insignificant. Across our eight sample sites in September, AI referrals produced around 430 human visits. As a channel, that looks like a failure.
But AI isn't a channel; it doesn't feed the funnel in the same way that search referrals do. It is an intermediary. And the moment you stop asking "what does AI send us?" and start asking "who is reading us, and why?", the whole problem changes character. You stop optimising a channel. You start serving an audience.
You have mapped two audiences. A third has arrived.
Every organisation we work with can describe its human audiences: the pension trustees, the in-house counsel, the procurement directors who decide whether to hire them. Most can describe their second audience too, the search engines, even if they outsource the relationship to an SEO agency. Audience thinking is not new. It is how marketing works.
What is new is that a third audience is now on your site every day, and almost nobody has mapped it. It has members. They have different motivations. Each one follows a journey through your content. And each one can end up recommending you, or not, to the buyers who increasingly ask AI assistants before they ask people, which is what makes it an intermediary.
AI is the new intermediary
Intermediaries are an audience you already know how to serve: the brokers, advisers and referrers who never buy from you themselves but decide who gets recommended, and who you probably market to as an audience in their own right precisely because of that power.
The machines are the newest of them and the fastest growing. They are not the buyers. But they read, they learn, and they pass recommendations to the people who are. Everything you know about marketing to an intermediary audience applies here: identify who matters, understand what job they are trying to get done, serve them the content that does it. The only new part is who the members are.
There is a fair objection to all of this, and we have had it put to us directly: the AI is not an audience, it is a layer. The customer is still the customer; the machine is the thing deciding which evidence gets carried into their decision. We are not disputing that: the buyer is still the buyer.
Why that matters
But marketers who sell through intermediaries already know that they need to be served differently. In financial services, it is the norm to sell mortgages through brokers and savings products directly to the customer. And the two audiences get very different content: the broker gets the full product range and all the background data, the customer gets the simpler version they can understand and act on.
Savings never had a broker. The customer was left to compare the products themselves, so the content was written for them; no one wrote the broker's version. Now an AI assistant has stepped into that gap: people are asking it to compare savings products the way they would once have asked a broker about mortgages, and it does the job with the only content it has, the simplified consumer version.
It is acting as an intermediary without the information an intermediary needs. That is why we call the operators an audience. It is not a claim about who pays. It is a way of classifying a set of jobs-to-be-done that are different from the human audience's, and serving them separately. Describing AI as a "layer" gives you nothing to work on; describing it as an "audience" gives you a method.
Who is actually in the audience?
The audience is not "AI": nobody is read by "AI", any more than you are hired by "the market". The audience members are the operators, with Amazon, Meta and a few others at the edges:
- OpenAI
- Microsoft
- Anthropic
- Perplexity
Five or six names, which is less than most organisations' key account lists.
"AI" is not a single audience
And they are not one audience but several: on one client site in September, OpenAI put nearly half of its live reads into the insight library and a quarter into the home/about pages, while Anthropic put a quarter of its live reads into the careers pages. Each operator has its own profile, so each needs its own map.
And here is the structure that makes them mappable: each operator sends a team of bots, and each bot in the team is responsible for a stage of the journey.
OpenAI sends three distinct readers to your site:
- GPTBot - crawls to train the models: that is the operator learning you exist, and your place in the market.
- OAI-SearchBot - builds the search index behind ChatGPT's browsing: that is the operator deciding you are worth referencing.
- ChatGPT-User - fetches a page live because a real person just asked a question: that is the operator reaching for you at the moment of decision.
Perplexity and Anthropic send teams of their own, with gaps we come to later. Google folds its AI reading into the same crawler that runs search, which is exactly why our previous article argued your Google AI problem and your SEO are one job.
If you have run account-based marketing, this is a familiar picture. The operator is the account. Its bots are the buying committee, each with a different job, each at a different stage. And unlike any human buying committee, this one leaves a complete record of its behaviour in your server logs.
What this audience reads
The measure here is live reads: pages fetched because a person has just asked a question and the assistant decided that page was worth reading for the answer. It is the nearest thing we have to a trust signal in the logs. What it shows is not what you think, and it is probably not what you spent your content time and effort on.
Across every site in this set, the pattern repeats: this audience reads the pages that demonstrate what you know, and largely ignores the pages built to convert.
On one client site in September, the news and knowledge section drew one live read for every three human page views; the product pages, which carry the commercial weight of the site, drew one for every twenty-five. On another, the insight library took nearly half of all live reads, and the service pages that describe what the organisation sells took one in thirteen.
Ratio of AI reads : Human page views:
Article pages
Product pages
Your content strategy
This is the finding that should change how you think about your content strategy. The content most organisations treat as decoration, the insight programme, the technical explainers, the research, is the content this audience consumes. Your thought leadership is not a brand exercise that sits upstream of the funnel. For the machine audience, it is the product. Your content strategy and your AI citability are the same strategy.
The five stages of the AI audience journey
Five stages, one operator at a time
Each operator sends a team, and each bot in the team is responsible for a stage. OpenAI's team shown as the worked example.
The AI audience map
The operator learns you exist. Bulk reading into the model’s memory.
The operator decides you are referenceable. Crawled into the live search index.
A live fetch because a person just asked. Observed citability.
What the assistant says about you. Google publishes a count; nobody else does.
A human arrives from an answer. Proof a citation happened.
Four of the five stages are in your server logs today. The diagnostics that matter are the gaps.
The map has three dimensions: which operator is reading, which stage of its journey it is at, and what type of content it is reading. You look at them two at a time:
- Operator by stage - tells you who covers you and where each one stops.
- Operator by content type - tells you what each one thinks you are for.
- Content type by stage - tells you which of your content is learned and which is cited.
This is the structure your content plan should be built on.
Start with operator by stage.
Because each operator's team is staged, you can place every operator on a journey through your content. There are five distinct stages, and you can observe four of them directly.
- Train - the operator ingests you into its models' memory. This is GPTBot, ClaudeBot, Meta's crawlers: bulk reading, around the clock, that determines whether the machine knows you exist at all. You see it in your server logs.
- Index - the operator's search bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot) crawl you into the live index behind their assistants' browsing. This is the operator deciding you are referenceable: are you worth live-reading when a fresh answer needs a source? This is in your logs too, with one exception: Anthropic's assistant searches through Brave, so its index stage shows up as Brave's crawl, not under an Anthropic name.
- Read - a live fetch at the moment of a real question. ChatGPT-User pulls a specific page because a person just asked something your content can answer. This is the strongest signal in the whole journey, observed citability, and it is in your logs with an IP you can verify.
- Represent - what the assistant actually says about you, from memory or from what it just read. This is the one stage you mostly cannot observe. Google alone publishes an appearance count: Search Console now reports how often your pages appeared in AI Overviews and AI Mode, but no other operator publishes anything. For the rest, you can interrogate the assistants directly and treat every answer as a snapshot, but not as a score. Our previous article explains why the tools claiming to measure this stage are guessing.
- Refer - a human arrives from an answer. Referral traffic from chatgpt.com or perplexity.ai, in your logs and (undercounted) in GA4. Every one is proof a citation happened; most citations do not produce one, which is why this is a baseline and a trend, not a volume metric.
Train, index, read, represent, refer. Each is a stage in the map, and the diagnostics that matter are the gaps: an operator stuck at one stage, or an audience where only one operator moves through them at all.
The stage gap
The first is the stage gap: an operator stuck at "Train" that does not arrive at "Read". Our own site is the cautionary example. Over four months, Anthropic's training crawler read our pages fourteen times for every one that Claude's live agent, the bot that fetches a page because a person asked, reached for. OpenAI, for comparison, fetched our pages live nearly three times as often as it crawled them for training.
Anthropic has learned us, Brave has us indexed, but Claude's live agent still rarely reads us live. When we tested it ourselves, the reason was rarely a choice against us: either Claude answered from memory and never searched, or the search it ran did not return our page.
The live agent can only read what the search step hands it. Learned, not cited. Knowing which operators are stuck there, and on which content, is what turns "improve our AI presence" into a work plan.
Our own site, June to September 2026
Operators down the side, stages across the top. One grid that shows the whole machine audience and where each member is active.
Operator coverage scorecard
Every cell except "Represent" comes from your server logs, with one footnoted exception.
compoundpartners.co.uk, 1 June to 30 September 2026. Verified and plausible bot identity only. Small site, four months: read as a worked example of the grid, not a benchmark. Perplexity's live agent is shown as not verifiable: the traffic carrying its name does not come from the address ranges Perplexity publishes, so we do not count it. *Anthropic's assistant searches through Brave Search rather than its own crawler; Brave's index holds our pages (checked 8 October 2026), so the cell is present, with no Anthropic-named crawl in the logs.
The coverage gap
The second is the coverage gap: an audience of five or six operators where only one reads you in any volume. In September, OpenAI carried at least 75% of live reads on every site in this set, and 95% at one; Anthropic carried most of the rest.
That concentration is comfortable right up until an operator changes its behaviour: in July, at one client, Perplexity's crawler cut its daily reading by about 80% inside a single month, and the site's machine readership concentrated even further into ChatGPT as a result.
Audience members can leave. On our own site, Meta's training crawler read eighty-odd pages in June and has not been back since. If your organisation's entire machine presence is one operator's habit, you have no record of the others and no way of noticing when it changes.
Serving it, and measuring it like an audience
None of this asks you to learn a new discipline. It asks you to re-apply the one you already have. The same five questions we ask about any audience apply equally to the new one, and they need to be asked operator by operator, because the answers differ:
- Who are they - which operators read you at all, and which send their whole team rather than only a training crawler
- Why do they read you - what questions is your content answering when the live fetches arrive
- What do they read - which pages define you in their eyes, and whether that matches what you want to be known for
- Where are they in the journey - train, index, read, represent, refer, plotted operator by operator
- How do you serve and measure it - the content that answers real questions, measured with audience metrics rather than channel metrics
That last distinction is the practical one. Channel metrics for this audience barely register: visits, clicks, referrals, all small, and all undercounted in GA4.
Audience metrics tell you what is happening: operator coverage, each operator's position across the five stages (operators down the side, stages across the top), what they read, and the trend in humans arriving from answers, read as a direction rather than a volume. That set of measures is what we mean by citability, and it is all visible today from data you already own.
One thing to be clear about: you cannot choose which operators read you. There is no media plan for this audience. What you can do is make sure nothing is blocking the ones that have gone quiet, and make sure the content they reach for is the content you want to be known by.
The machines are already in the room
There is a symmetry here with ABM. You would never run a key account on guesswork, and you would never ignore an intermediary who influences your best referrals. You would map the committee, learn what each member cares about, and notice when one goes quiet.
Yet the intermediary audience that now reads some organisations' sites almost as much as their human visitors do is unmapped at almost every organisation we have looked at, while the effort goes into a visibility score nobody can be confident in.
The machines are already in the room, reading. The questions that matter are what they are reading, whether it says what you want it to say, and whether you would know if one of them stopped.
What next?
Map your most important audience:
If you want to know what one audience actually sees when it meets your organisation, our free Second Opinion assesses your organisation from the outside, through two lenses: how the human audience finds and reads you, and how the machines find, read and represent you.
Request a second opinion
One audience of your choice. To ensure high quality, we only work with two organisations each month.
Read your server logs
And if this article has raised a question your own server logs should answer, decoding them is where we would start.
Get a snapshot of your logfiles
FAQs
Because the machines behave like intermediaries, not like a traffic source. A channel sends you visitors; across eight client websites in September 2026, AI answers sent around 430 human visits while AI assistants fetched 26,500 pages live. The assistants read your content, learn from it and pass recommendations to buyers who ask them before they ask people, which is what brokers, advisers and referrers have always done. Organisations already know how to serve an intermediary audience: work out who matters, what job they are trying to do, and what content does that job.
Not "AI" but the operators: OpenAI, Google, Microsoft, Anthropic and Perplexity, with Amazon, Meta and a few others at the edges. Five or six names, fewer than most key account lists. Each sends a team of bots, and each bot is responsible for a stage: a training crawler that learns you exist, a search bot that decides you are referenceable, and a live agent that fetches a page because a person just asked a question. OpenAI's team is GPTBot, OAI-SearchBot and ChatGPT-User.
A page fetched by an AI assistant at the moment of a real question, because the assistant decided that page was worth reading for the answer. It appears in your server logs as ChatGPT-User, Claude-User or Perplexity-User, with an IP address you can verify against the operator's published list. It is the nearest thing to a trust signal the logs contain, and it is the measure this article uses. Training and search-index crawls are stages on each operator's journey, but they rise and fall for reasons you cannot see from the outside, so they are not counted as reads.
The pages that demonstrate what you know, not the pages built to convert. On one client site in September 2026 the news and knowledge section drew one live AI read for every three human page views; the product pages drew one for every twenty-five. On another, the insight library took nearly half of all live reads and the service pages one in thirteen. For the machine audience, thought leadership is not upstream of the funnel; it is the product.
Train, index, read, represent, refer. Train is the operator ingesting your content into its models' memory; index is its search bot crawling you into the live index behind the assistant; read is a live fetch at the moment of a question; represent is what the assistant says about you; refer is a human arriving from an answer. Four of the five are observable in your own server logs. Represent is the exception: only Google publishes an appearance count, through Search Console.
With audience metrics rather than channel metrics. Channel metrics (visits, clicks, referrals) are small and undercounted. Audience metrics are operator coverage, each operator's position across the five stages, what each one reads, and the trend in humans arriving from answers, read as a direction rather than a volume. That set of measures is what we call citability, and all of it is visible today from data you already own. You cannot choose which operators read you, but you can make sure nothing is blocking the quiet ones and that the content they reach for is the content you want to be known by.
Want more?
Sign up for our Second Order Thinking newsletter
Second order thinking
It’s not just what happens, it’s what happens next that matters.
-
Stop calling it "AI visibility"
Second order thinking
Ask three marketers what "AI visibility" means and you will get three different answers, none of them measurable. Using more th…
-
Your website has three jobs now. Most professional services organisations are only doing one
Second order thinking
Over half of Google searches now end without a click. That stat should change how every professional services organisation thin…
-
Measuring brand visibility in 2026: what you’re missing and why it matters
Second order thinking
AI-driven answers are impacting your brand visibility and traditional measurement no longer shows the full impact.