AI Search Rewards What the Web Can’t Easily Summarise

Every content brief we read now carries a version of the same instruction. Publish original content, because AI can summarise everything else. It is correct, and it is nearly useless, because almost nobody says what original means.

The advice usually collapses into three suggestions: run a survey, quote an in-house expert, add a case study. All three can be done badly enough to produce something a language model could have written unaided. A survey of two hundred marketers about their top challenges. An expert quote that restates the paragraph above it. A case study with the client’s name removed and the numbers rounded until they say nothing. That material is original in the sense that nobody published it first. It is not original in the sense that matters.

The test is not novelty. The test is whether somebody else could reproduce it.

Could a stranger make a fair copy of this without you?

Ask it plainly. Could a competent writer with a search engine and a good model produce a fair equivalent of this page in an afternoon, with no access to anything of yours?

If yes, the page is commodity information regardless of who typed it. It may still be worth publishing — people need orientation, and a clear explainer earns trust. But it is not an asset. It is inventory anyone can restock.

If no, the next question is why not, and the answer is always the same shape. The information exists because your organisation did something. Somebody watched a system fail. Somebody chose one approach over another and lived with the consequences. Somebody measured a thing nobody had measured. Somebody made a call where the evidence did not settle it.

That is the whole thesis. Durable web content is the record of costly acts. Everything else compresses.

Google’s own documentation points the same way. Its guidance on generative AI features asks publishers not to recycle what others have said or what a model could easily produce, and contrasts commodity pages with unique expert or experienced takes. Its quality rater guidelines go further, telling raters to weigh effort, originality and skill, and to give the lowest rating to content produced with little of any — whether a person or a machine made it. Effort is not a moral test there. It is a proxy for exactly this: something had to happen for the page to exist.

Six kinds of information that do not compress

A practitioner’s list, not a standard.

First-party observation. What you saw, from a vantage point only you occupy. Pew Research Center could report in July 2025 that people clicked a result link on 8 per cent of visits where an AI summary appeared, against 15 per cent where none did, because it had recruited 900 US adults who agreed to share their browsing. Nobody can synthesise that number. It did not exist until somebody paid for it to. Note what Pew claimed, too: an association, not a cause. Report your own observations the same way.

Proprietary process. Not a five-step diagram. The actual sequence you follow, why it runs in that order, and what you refuse to do. Our own refusal list is public: no media buying at scale, no PR, no event production. A model can describe a discovery process in general. It cannot say why yours is two weeks and fixed-fee, or what you learned the first time it was not.

Original analysis. A claim you derived yourself, with the method shown. The method is the load-bearing part. An unexplained number is a rumour, and rumours compress fine.

Documented decisions. The most undervalued category, because the value is in the road not taken — the option you rejected and the constraint that killed it. We have led frontend architecture for an industrial platform where the alerting layer had to combine several kinds of logic, and the hard question was never how to calculate a threshold. It was which alerts an operator would act on, and which ones trained them to ignore the screen. No amount of general knowledge produces that answer for a specific plant.

Evidence. Artefacts with a method and a date attached: an audit, a test result, a screenshot of the real tool. Undated evidence decays into assertion.

Expert judgment under uncertainty. A call made where the evidence did not settle it, with the uncertainty left visible. Models are structurally weak here, because they average the published consensus. Shipping one product to five television app stores taught us that the certification calendar, not the code, is the schedule risk — a judgment that looks obvious afterwards and is worth real money in front of a roadmap.

Most companies have no research to publish

This is the strongest objection, and it is usually right. A forty-person firm does not have a research function. It has a delivery schedule. Telling it to publish original data is telling it to become a different company.

But research and originality are not the same thing. Every operating business generates a record whether or not anyone writes it down: the question that comes up in every sales conversation, the requirement always underestimated, the failure that recurs across clients, the reason you priced something the way you did. The constraint is not research capacity. It is documentation discipline. The material already exists — it lives in people’s heads and in Slack, and it leaves when they do.

Two concessions. Some of the best material is genuinely confidential, and de-identifying it well is real editorial work rather than find-and-replace. And a few firms really do have nothing publishable yet, usually new ones or ones whose work is entirely someone else’s. Those firms should publish less, not more.

An inventory you can run this week

Take a room, two hours, and the three or four people who actually do the work. Not the marketing team alone.

1. List the last twenty questions a prospect or client asked. Straight from inboxes and call notes. Do not tidy them.
2. Mark each one C or O. C if a competent stranger could answer it well with a search engine. O if answering it properly needs something of yours — a number, a case, a decision, a scar.
3. For every O, name the source. Which of the six categories does it come from, and where does the raw material live? If nobody can name the source, it is an opinion you have not earned yet.
4. Check what confidentiality costs. What has to come out for this to be publishable, and does it survive the removal? If not, drop it. A story stripped of everything specific is worse than silence.
5. Rank by the gap, not by search volume. The best item is the one where the distance between what you know and what the open web knows is widest — even if almost nobody searches for it. Volume is what the growth work is for.
6. Assign an author who was there. This is why our own Insights are written by the people on the project, and published when there is something to say.

The ratio is the finding. We are not going to tell you what it usually comes to — we have run this often enough to have an impression and not often enough to have a number, and a benchmark invented for rhetorical convenience is exactly the move this article argues against. What matters is your ratio, and that teams are generally surprised in the same direction.

The O column is the part worth defending and expanding. Making it findable is the next problem, and a smaller one than most teams assume. When we are asked to help with AI and automation work, this inventory usually comes before anything gets built — because a system that surfaces nothing you uniquely know is an expensive way to repeat the internet back to your customers.