Your website now has two audiences. You see one.
Until 2022, the only audience for your website was the people who visited it. Now there's a second one: the language models that read your pages without ever rendering them. The second audience already outnumbers the first on most sites — and they want different things.
The shift that happened quietly
Search-engine bots have always read websites. Googlebot has been crawling and indexing pages since 1998. What changed isn't that bots exist — it's what bots do with the content they read.
Googlebot's job has historically been to send people to pages. The bot reads the content, ranks it, and shows it in search results so a human can click through. The bot is a librarian, not a reader. It cares whether your page exists and what it's about, but it doesn't try to extract the information itself.
The new bots — GPTBot, ClaudeBot, PerplexityBot, Google-Extended for AI training, and the AI Overviews crawler — read pages so that an AI assistant can answer a question without sending the user to your page at all. They're not librarians. They're readers. They extract information, paraphrase it, and present it to the user as a synthesised answer. Your page is the source. Your visitor count from these audiences is often zero.
This is the shift. The audience that reads your site now outnumbers the audience that clicks on it. By analytics, you only see one. By citation footprint, the other is doing most of the work.
Why optimising for one increasingly optimises against the other
If both audiences wanted the same things, this wouldn't matter — write for humans and AI follows. That was largely true until 2022. It's no longer true now, because the two audiences have diverged on at least seven dimensions:
Concise prose. Get to the point. Cut the throat-clearing.
Comprehensive coverage. More context = more potential citation hooks.
Personality. Voice. A reason to trust this writer specifically.
Neutral, factual framing. Personality language tends to get stripped during summarisation.
Clear CTAs. Buy now. Get a quote. Book a call.
To ignore CTAs entirely. Action language is filtered out during retrieval as non-substantive.
Visuals. Images, diagrams, video.
To read alt text and ignore the image itself. Most AI crawlers don't render images. Some don't render JavaScript either.
Storytelling and analogy. Make abstract ideas concrete.
Direct claims that can be extracted as facts. Analogy gets paraphrased away.
Visual hierarchy: headings, white space, scannable structure.
The same thing, mostly. This is the one dimension where the audiences mostly agree.
Navigation that helps them find related content.
To ignore nav menus during content extraction. Internal links are read for site graph purposes but not for the page's content.
The wrong way to handle this: cloaking
The naive solution is to serve different content to bots than to humans — render a clean, structured, machine-readable version when the user agent is GPTBot, and the regular site when the user agent is a browser. This is called cloaking and it has been a Google policy violation since the early 2000s. The penalty for getting caught is severe and the AI crawlers have learned to spot it. Don't do this.
The right way: one page, layered to serve both
The solution most agencies miss is that you can serve both audiences from one page if you think carefully about what each piece of the page does.
Layer 1 — the human surface
This is what visitors see. Hero image, punchy headline, short subhead, three benefit bullets, a CTA. Designed for skim-reading and conversion. Write this for humans first.
Layer 2 — the substantive body
Below the hero, write the content the page is actually about. Direct claims. Specific facts. Named entities. Verifiable numbers. This layer is where both audiences live: humans read it if they're qualified buyers who care, and bots extract from it for citation. Write to be useful to a reader who already wants to know — and the bots will treat that usefulness as extractable signal.
Layer 3 — the structured data
Embed schema markup that describes what's on the page. Article schema for blog posts. LocalBusiness for service pages. Person for author pages. The schema doesn't show up on the page; it shows up in the source code that bots read. This is the cleanest channel for telling AI assistants exactly what your page is, who wrote it, and how it relates to other entities.
Layer 4 — the supporting evidence
Outbound citations, internal links, named sources, footnotes if appropriate. Humans skim past these unless they're motivated to verify. Bots read them as authority signals — the more verifiable references a page contains, the more confident the AI is in citing it.
The page that serves both audiences well has all four layers. Most pages have only the first.
In my 2021 systematic review of chatbots in digital business transformation (IEEE Access, 148 citations), my co-authors and I identified the "human-in-the-loop" pattern as the recurring design that makes AI systems trustworthy in business contexts. The paper was about chatbots, but the underlying principle generalises: AI systems work best when human and machine consumption are designed together, not as separate optimisation targets. The two-audience framework for websites is the content equivalent of that pattern — design for both readers in one artefact, not for either reader in isolation.
What this looks like in practice
A practical example. Imagine a Melbourne accounting firm's services page. The old single-audience version:
"Our experienced team delivers tailored accounting solutions for growing businesses. Get in touch today to discuss your needs!"
Pure layer 1. Humans skim it. Bots find no extractable substance.
The two-audience version:
"Tax and compliance for Melbourne-based businesses between $1M and $20M revenue. Founded in 2014. Currently working with 87 clients across professional services, hospitality, and trades. Recognised by CPA Australia, recommended by the Victorian Chamber of Commerce."
Same length. Vastly different signal density. The human gets the same scannable summary, but now with specific information they can use to decide if you fit them. The bot gets a list of extractable facts: founding year, geo, revenue range, vertical, client count, credentials, recognition. Both audiences served. No cloaking.
Where to draw the line
The temptation, once you understand the framework, is to over-optimise for the bot side. Stuff the page with extractable facts. Maximise entity references. Cite every claim. Resist this.
Bots are now sophisticated enough to identify boilerplate. A page that reads like it was written for AI extraction will get classified as low-quality content and lose citation weight. The same overstuffing penalty that hurt keyword SEO in 2012 will hurt entity-stuffing in 2026 — except faster.
The line to draw: write content that is genuinely useful to a human who already wants to know what you're saying, and then add structural support (schema, citations, named entities, specific facts) that makes that usefulness verifiable. The order matters. Human usefulness first, bot extractability second. Sites that do it the other way around end up with content that serves neither audience well.
Want to see which audience your current site serves better?
Our free AI search audit walks your homepage and key service pages through both lenses — what humans see and what bots extract — and gives you a three-page report with the specific gaps and where each audience is being under-served. AU SMBs only. No upsell.
The bottom line
Your website has two audiences whether you've designed for them or not. The bot audience is reading your content, deciding whether to cite you, and increasingly answering user queries on your behalf. The human audience is the smaller, more downstream consumer of all of this. If you only think about the human audience, the bot audience will under-cite you and you'll lose visibility you didn't even know you had access to. If you only think about the bot audience, your human visitors will bounce.
The work isn't choosing between them. The work is designing pages where both audiences find what they need from the same content, in the right layers, with the right specificity.
For the practical follow-on, see the 11 words that make ChatGPT recommend you — that piece is the granular implementation of the framework: specific phrases that satisfy both audiences simultaneously.