Search stopped sending the click. Build the thing it cannot summarise.

Section 01What actually changed

The web's founding commercial bargain was simple. You published an answer, a search engine indexed it, and when someone asked the question you received a visit. Content was the toll you paid for traffic. That bargain is being unwound, and the numbers are no longer ambiguous.

That figure is the headline, but it is the second-order numbers that should change how you allocate a budget. When Google shows an AI summary, the behaviour of the person reading it changes sharply and measurably.

And the surface is expanding, not stabilising. The share of Google keywords triggering an AI Overview grew from roughly 1.5% to about 32% between September 2024 and September 2025 — a twentyfold increase in twelve months. Whatever proportion of your queries carry an AI answer today, the planning assumption should be that it is materially higher in a year.

Read together, these describe a specific change rather than a general apocalypse. The engines have not stopped needing content. They have stopped needing to send you traffic in order to use it. The question for anyone whose business depends on being found is no longer how to rank. It is what to publish that still requires the visit.

Section 02Why both standard responses are wrong

Two positions dominate the conversation, and neither survives contact with the data.

"SEO is dead"

It is not. Classic ranking signals still gate the candidate pool that retrieval draws from. A page that cannot rank cannot be retrieved, and a domain with no authority does not get considered in the first place. What has died is a specific, narrow trade: publishing a competent general answer to a factual question and being paid in clicks. That was always the most commoditised end of the market, and it is the end that a language model replicates most cheaply.

"Just do AEO"

The bigger problem. An entire advisory category has appeared in under two years, and a great deal of what it sells is untested. Three of its most repeated recommendations range from unsupported to actively contradicted — covered in full in the next section. Buying a checklist whose items have never been isolated as variables is not a strategy; it is a hedge purchased from someone who benefits from your anxiety.

The useful response is narrower and less saleable than either. Work out which of your pages exist to deliver a fact — those are the ones being taken — and replace their commercial job with something a summary cannot do.

Section 03What the evidence supports — and what is oversold

There is one piece of real academic work underneath most of this field, and it is worth knowing precisely what it says and how old it is.

Freshness holds up

Three things being oversold

Strip out the folklore and a short, unglamorous list remains: be reachable, be structured as answers, carry evidence worth quoting, be a clear entity, and be corroborated by sources you do not own. That is the whole of it. It is not proprietary and it is not complicated. It is just work.

Section 04Why tools survive, and the reason it is not the one you were sold

The real argument does not depend on engine behaviour at all. It depends on what the query is for.

An AI answer can replace a fact. It cannot replace a computation performed on the reader's own numbers. "What is a normal day rate for a copywriter" is a fact, and a model will answer it completely, and you will never see that person. "What should I charge" is not a fact — it depends on their take-home target, their business costs, their billable ratio and their tax position, none of which the model has. The only way to answer it is to run it.

This is a structural distinction, not a tactical one. The commercial value of a page is now roughly inverse to how completely a summary can substitute for visiting it. Ranked from most exposed to least:

There is a second effect worth naming, and it is the part most people miss. A tool page has two commercial jobs, and they do not compete. The tool captures the visit. The prose around it — the methodology, the worked example, the honest caveat, the published range — is exactly the citable, statistic-bearing, source-citing material the GEO research found performs best. Building the tool forces you to write the thing most likely to be quoted.

Section 05The twelve archetypes

Nearly every useful micro-tool is one of twelve shapes. The value of knowing the list is that you can run it against any market and generate a build queue in an afternoon — the archetype tells you what the tool does, and the market tells you what it is about.

The twelve tools published alongside this report are one of each, deliberately. The point of the set is not the individual tools; it is that the shapes transfer. A dental practice, a med spa, a law firm and a SaaS company need completely different tools and exactly the same twelve archetypes.

Section 06Where Google's line actually is

The obvious objection: if tools are this good and this cheap to build, does mass-producing them get you penalised? The policy language is clearer than most people assume.

The load-bearing word is primary purpose. The method is explicitly not the violation. The absence of user value at scale is. Which gives a practical test that is easy to apply and hard to fake:

• Can a visitor complete a real task on this page without leaving it?
• Does the page contain information the reader could not get elsewhere — your data, your method, your ranges?
• Would you show it to a client without apologising for it?
• Would you have built it if search did not exist?

A tool that passes all four is not what the policy is aimed at, however many of them you build. A page that fails the first — a templated shell with a thin formula, a keyword in the title and no completable task — is precisely the pattern that got sites deindexed in 2024, and the enforcement cadence has got faster since, not slower.

Section 07The economics after AI-assisted build

Until recently a calculator page meant developer time — a specification, a sprint, QA, a few thousand pounds and a queue. That cost has collapsed. A competent operator with an AI coding assistant can now produce a genuinely good tool in hours.

The naive conclusion is that this makes the strategy better. It does the opposite, and understanding why is the whole game: when production cost collapses, production stops being the moat. Everyone can now build the template. What did not get cheaper is the part that was always scarce.

So the correct response to cheap production is not more tools. It is better ones: a defensible model, numbers only you have, a published methodology, and a level of finish that signals somebody cared. Those were always the differentiators. AI simply removed the excuse for not clearing the bar.

Section 08Distribution: how a tool gets found

Tool pages attract links at a higher rate than blog posts, and the mechanism is worth stating precisely because it dictates the tactics. A blog post gets cited when someone agrees with it. A tool gets referenced when someone needs it — and unlike a "best of 2024" post, it does not go stale, so it accumulates for years rather than decaying.

Five channels actually do the work:

• Round-up placement. "Best X calculators" lists already rank for your term and are already shaped like the answer an assistant wants. Being worth listing, then asking, is unglamorous and effective.
• Embeds. An embeddable version placed on partner sites is both a link and a lead surface on somebody else's traffic. This is how mortgage calculators colonised the property web.
• Communities. Analysis of a large prompt sample found ChatGPT leans heavily on Reddit. A tool that genuinely answers a recurring question in a subreddit gets linked by people who are not you. Astroturfing is detectable and expensive.
• The result as the share unit. Every result should have a permalink encoding its inputs. People do not share tools; they share their number, and the tool travels with it.
• Directories and launch platforms. Real reach, but note that Product Hunt links are nofollow by default — treat it as audience, not link equity.

Section 09Selling this: two very different markets

The same capability sells into two markets that want almost opposite things, and conflating them is the most common way this goes wrong.

Market one: the business that needs to be found

A dental practice, a med spa, a law firm, a clinic. They do not want a search strategy. They want more of the right enquiries, and they are being told by everyone that AI is about to take their traffic. Sell them a specific asset with a specific job: the page that answers "how much do implants cost" so well that the answer engine quotes you and the patient still comes to you to run their own numbers.

Price it as a discrete deliverable, not a retainer. It is a durable asset with a fixed scope, which is far easier to approve than an open-ended monthly fee — and it produces something the owner can see, which retainer content usually does not.

Market two: the consultant who serves them

Freelancers, copywriters, agency owners and fractional operators — people already connected to several brands and actively looking for something new to sell. They do not need to be convinced the market exists. They need a repeatable thing they can take to clients they already have.

What they are buying is the method: the twelve archetypes, the question research that decides which one, the model that makes it defensible, and the build standard that makes it look like a product rather than a form. One consultant with five clients has five markets and a build queue.

Section 10What makes a tool fail

Every failure mode below was observed in the research rather than invented for symmetry.

• It is a template with no unique input. A common formula in a new wrapper is the textbook scaled-content pattern, and it is what the policy is actually aimed at.
• Volume outruns authority. Hundreds of pages on a domain that has earned little trust is an attack surface, not a strategy.
• It targets a head term. "Mortgage calculator" is a keyword difficulty of 100 owned by Bankrate, NerdWallet and Zillow. The winnable ground is specific: a niche, a region, a segment, a combination nobody has bothered with.
• The result is not the point. If the real deliverable is an email capture and the number is bait, people notice immediately and it never gets linked.
• There is no share unit. No permalink, no copyable summary, no export. The tool works and then evaporates.
• It renders only in JavaScript. Nothing to retrieve, nothing to cite. See section 04.
• Nobody checked the maths. One wrong number in public, found by one person who knows the domain, costs more trust than the tool was ever going to earn.
• It never gets touched again. Freshness is a real signal and enforcement has got faster. A tool is a small product, not a campaign.

The through-line: every one of these is a decision to do less work than the thing required. Which is the reassuring part of the whole analysis. The strategy is not clever and it is not secret. It is just more work than most people will do, on a bar that AI raised rather than lowered.