Thinking · 18 September 2026

Is your site answer-ready?

A named standard, six checks you can run yourself in ten minutes, and the three reasons most pages fail them.

A page is answer-ready when an AI can lift the answer straight from it: the claim is stated plainly, in text, early, and the page carries enough context to stand on its own without the rest of the site.

It is a page-level standard, not a site-level one. A site can be beautifully built and still have not one answer-ready page on it, which is the usual situation.

The six checks

Take one page, ideally the one you would most like to be recommended for, and run these. Each is pass or fail, and ten minutes is enough for all six.

1. Is the answer in text?

Not in an image, a PDF, a slider or a video. If the only place your price appears is a downloadable rate card, the page does not contain the answer.

Fails: a pricing table rendered as a graphic; a tariff in a PDF; delivery areas shown only on an embedded map. Passes: the same information as text, with the image kept for the people who prefer to look rather than read.

2. Is it in the first screen of content?

An answer buried under four paragraphs of positioning reads, to anything scanning for the claim, like a page that does not make one. Position is a proxy for importance, and burying the answer signals that the answer is not the point of the page.

3. Is it stated, not implied?

“Competitive pricing” is not a price. “Fast turnaround” is not a timescale. “Enterprise-grade” is not a specification. A claim a reader has to infer is a claim that cannot be quoted, and a claim that cannot be quoted cannot be attributed to you.

The test is whether the sentence would survive being extracted on its own. “We offer competitive pricing” extracted on its own says nothing. “One-page sites are £1,200 and take two days” survives extraction intact.

4. Does the page stand alone?

If understanding it requires having read two other pages first, then quoting it in isolation produces something misleading, and isolation is exactly how it will be quoted. A model does not arrive at your page having read your navigation in order.

This is the check that most often argues against tidy information architecture, and the resolution is deliberate repetition. If the price only makes sense alongside what is included, then what is included belongs on the pricing page too, even though it also lives on the services page.

5. Is it self-describing?

Does the page say what it is about near the top, in the words a person would actually use for it? A heading that only makes sense once you have read the page is a heading that cannot be matched to a question.

“Built different” describes nothing. “What a website costs, and what is included at each price” describes exactly one thing, and matches what someone would actually ask.

6. Is it marked up, and does the markup agree?

Structured data that matches what the visible text says. Parity matters more than presence: schema that contradicts the page is worse than no schema, because it makes the page unreliable rather than merely unhelpful. A price of £1,200 in the markup and “from £900” in the copy is not a technicality; it is a reason to trust neither.

The seventh check, which overrides the other six

A page can pass all six and still be invisible, and this is the failure most likely to be news.

Vercel and MERJ analysed a month of AI crawler traffic across Vercel’s network (569 million requests from GPTBot, 370 million from ClaudeBot) and found that none of the major AI crawlers execute JavaScript. They fetch JavaScript files and read them as text without ever running them. Googlebot renders JavaScript; the AI crawlers do not.

So if your page builds its content in the browser, everything above is irrelevant, because none of it is present when the machine reads the page. The answer can be perfectly stated, early, in text, and still arrive as an empty container.

How to check properly: view the raw page source rather than the browser inspector. The inspector shows the page after JavaScript has run, which is why this problem survives so long inside teams: everyone keeps looking at the rendered version and it looks fine. If you would rather not read HTML, disable JavaScript in your browser and reload. What remains is approximately what an assistant gets.

Why most pages fail

Three causes, and none of them is laziness.

Throat-clearing

Good writing often builds to its point. Answer-ready writing leads with it and builds afterwards. These are genuinely in tension, and the resolution is not to write badly: state the claim, then earn it in the paragraphs below. The page still reads well to a human, and the answer is where a machine will look.

This is the objection worth taking seriously, because it usually comes from whoever cares most about the writing. The honest answer is that leading with the claim is a constraint, like writing to a word count, and constraints tend to improve marketing copy rather than damage it. A page that cannot state its own point in its first sentence is often a page whose author had not settled what the point was.

Context spread across the site

Site architecture encourages splitting a subject across several pages, each covering its part. It makes a tidy navigation and an unquotable set of pages, because no single one of them contains a complete answer. Some deliberate repetition between pages is the cost of each one standing alone, and it is worth paying.

Answers deliberately withheld

The commonest and most deliberate failure: the answer is missing because somebody decided it should be, usually a price. That is a real commercial choice with real reasons, but it should be made knowingly. A page that withholds its answer cannot be the source of that answer, and something else will be. We publish ours for that reason, among others.

A worked example

A services page opening with “We partner with ambitious brands to deliver transformative digital experiences” fails checks 2, 3 and 5: the claim is not in the first screen, nothing is stated, and the page does not say what it is about in words anyone searches with.

The same page opening with “We build websites for UK businesses. Four fixed prices, from £1,200 for a one-page site to £60,000 for a complex build, delivered in two days to five weeks” passes all three, and has not become less persuasive for it. The positioning still works; it is just underneath the answer rather than in front of it.

Note what did not change: the tone, the design, the brand. The rewrite is one paragraph moved and three facts made explicit. Answer-ready is not a redesign, and treating it as one is how it ends up on a roadmap instead of being done this week.

When you genuinely cannot publish the answer

Some answers cannot go on a page, and pretending otherwise is not useful. Regulated pricing, individually negotiated contracts, anything covered by a client NDA, anything where the real answer is “it depends” in a way that is true rather than evasive.

The move is not to leave the question unanswered. It is to answer the nearest question you can answer honestly, and say why you are answering that one. A range instead of a price. A typical case instead of a quote. The three factors that move the number, named, so a reader can place themselves. “Most projects of this type land between X and Y; what moves it is A, B and C” is quotable, checkable and true, and it is worth a great deal more than “contact us for a quote”.

The test of whether a withheld answer is a real constraint or a habit: ask whether you would say it out loud on a first call. Almost everyone would give a range on the phone within five minutes. If you would say it to one person, the only thing publishing changes is how many people have to ask before they hear it.

Testing it the way a machine would

Three checks, in ascending order of effort.

1. Read the raw source Search it for your headline and your price. If they are not there as text, stop here and fix that first.
2. Paste the page text into an assistant and ask it your buyer’s question Not “summarise this”. Ask the actual question, such as what this costs or how long it takes. If it cannot answer from the page alone, the page is not answer-ready, and you have just watched it fail in the same way it will fail for real.
3. Ask an assistant the same question without the page and see who it names instead. That tells you who currently owns the answer you are trying to own.

What to do with the result

Fix the page you most want to be recommended for first, then the two behind it. This is not a project. Most pages need a rewritten opening paragraph, one price or timescale made explicit, and schema that agrees with the text. The structural work (server-side rendering, entity resolution, a machine-readable source of truth) sits at the site level and is covered under AI-native.

Get a page reviewed The glossary

Sources

• Vercel and MERJ, The rise of the AI crawler: one month of crawler traffic across Vercel’s network, including 569 million GPTBot requests and 370 million ClaudeBot requests.
• Jeremy Howard, llmstxt.org: the llms.txt proposal, first published September 2024, second version August 2026.