~1%
Current share
Of web traffic comes from AI assistants today.
AI referral traffic
How much traffic do assistants actually send?
About one per cent of the web. Anyone telling you the game has already changed is selling something. The interesting part is the shape of the curve, not its height, and the visitors who do arrive come pre-qualified, because they asked their questions before they clicked.
10x
Referral growth
Growth in AI referral sessions since late 2024.
4x+
Conversion
Conversion against organic search.
The sceptical view
Isn't this just SEO with a new invoice?
Often, yes. Much of what sells as answer engine optimisation is SEO fundamentals repackaged at a premium, and major publishers still see under one per cent of their referrals from AI platforms despite being cited constantly. The sceptics have the hype exactly right.
Where they lose the argument is direction. Agents no longer just read; they act. ChatGPT handles some fifty million shopping queries a day, Stripe and OpenAI have shipped a protocol for agent-led checkout, and Google is building agentic browsing into Chrome itself. The web is getting an agent lane. A site that machines cannot read is not on the road.
Published source
"Companies will need to focus on producing unique content that is useful to customers... expertise, experience, authoritativeness and trustworthiness."
Alan Antin · Gartner, on the coming 25% decline in traditional search volume
Content structure
Why can't a legacy CMS serve a machine reader?
Because it stores pictures of pages rather than content. On an older platform the words live inside WYSIWYG fragments with layout baked in, the markup is locked inside templates written a decade ago, and there is no structured data because nothing structured exists to express. A human squints past all of this. An agent does not squint. It fails quietly and quotes a competitor whose content it could parse.
How content is modelled inside the CMS now decides whether machines can read it at all. Structured content used to be an engineering preference. It is becoming distribution.
Two readers
Layout and meaning are locked together in WYSIWYG content.
The reader cannot reliably extract the facts.
The same content is explicit, attributable and reusable.
The machine can understand and cite the source.
People read the design; machines depend on the page structure and metadata.
Start with the structure
What should you change first?
01
Question-shaped headings
Headings that state the question a reader would ask, because that is the string an assistant matches against.
02
Answer-first openings
The answer in the first paragraph, complete enough to quote. Machines lift openings; so do busy people.
03
Valid structured data
Facts stated in JSON-LD as well as prose: what the organisation is, what the page answers, what the numbers are.
04
One identity everywhere
The same name, claims and numbers on every page and profile. Assistants cross-check, and inconsistency reads as unreliability.
05
A maintained llms.txt
A plain-text introduction to the organisation at the root of the site, kept current, so the assistant starts from your words.
TEST
Then the free test
Ask three assistants what your company does and who they would recommend in your category. The gap between their answer and the one you would want is the work.
The published standard
Quotable, consistent and machine-legible.
All five changes come from our published Scrape-Bias Page Standard. Nobody can buy placement in an AI answer. What you control is being quotable, consistent and machine-legible, and that is engineering.
Asked plainly
Frequently asked questions.
How much traffic do AI assistants send to websites?
Around 1% of web traffic in 2026, growing at roughly a percentage point a month, with sessions up nearly tenfold since late 2024. The visitors they refer convert at several times the rate of organic search.
Is answer engine optimisation just SEO rebranded?
Much of what is sold under the label is repackaged SEO fundamentals, and the sceptics are right to say so. The underlying shift is real: content must now be legible to machine readers as well as people, and that is decided by structure, not marketing.
Why do legacy CMS platforms struggle with AI assistants?
Because they store content as visual page fragments rather than structured data. An assistant cannot reliably parse WYSIWYG markup, template-locked layouts or pages without structured data, so it quotes a clearer source instead.
What should we change first?
Five things: question-shaped headings, answer-first opening paragraphs, valid structured data, one consistent identity everywhere, and a maintained llms.txt file at the root of the site.
Sources
Claims checked July 2026.
Next step
Make your platform legible.
The page standard is published and free to adopt. If you would rather see how your own site reads to a machine, that is a short conversation with an engineer.