You write out your opening hours, your price range and your common questions clearly. The better you write them, the more readily an AI reads them aloud to the user, and the less reason that user has to click through.

It sounds like a penalty for doing the job well. What it’s really telling you is narrower: your site needs to hold something an answer can’t carry away.

A page has three layers, and most people build two

Take a content page apart and it’s serving two readers at once — a machine collecting answers, and a person making a decision. The sensible arrangement has three layers.

The answer layer holds the conclusion. Two to four sentences under the heading that answer the question directly, with a complete subject, the conditions attached, and a date. This layer exists to be lifted by an AI, and to serve readers who only want the result.

The evidence layer holds the backing. Source data, citations, comparison tables, worked examples, how you arrived at the number. This layer decides whether an AI trusts you, and whether a person keeps reading.

The conversion layer holds the next step. This is the “so what now” — and it has to be something that can only be completed on your site.

We’ve written about the first two, and most people execute them acceptably. The failure rate on the third is much higher, because in the search era it barely mattered: the visitor was already on your page, so a contact form was enough.

The third layer is now the one that decides whether you have a business

Search-era logic ran ranking → click → conversion. Three segments joined end to end; get the first right and the rest followed.

The join between the first two has broken. When an AI Overview appears, the top-ranking page’s average clickthrough rate drops 58% (Ahrefs, February 2026), and the share of AI Overview citations coming from the top 10 has fallen from 76% to 38%. However well you write the answer layer, what you win is being read aloud, not being clicked through to.

Leaving a “contact us” form as your third layer at this point means betting the whole funnel on the 42% of clicks that weren’t intercepted.

The conversion layer needs to become something else: not an invitation to come in and contact you, but a reason it’s impossible not to.

What AI can’t carry away is anything that needs YOUR data

A summary can copy static text. But “the user has to do something themselves” has stopped being a barrier.

One variable changes the judgement here: Google’s Generative UI, which until recently ran only in AI Mode, is now reaching AI Overviews. It builds an interactive interface on the spot to match the question — mortgage calculators included. So building a tool that takes an input is no longer a moat on its own. Anything whose formula is public, and which needs none of your data to compute, the AI now generates inside the answer, where the user never has to leave.

The line has moved back a step, to where the inputs come from.

First: calculations that need your data to run. Not monthly loan repayments, whose formula anyone can look up — your actual pricing bands, your lead times and surcharge rules, the specs only you have measured. Generative UI can build the shell of a calculator. It can’t supply your numbers.

Second: filtering and comparison. Hundreds of specifications, models, listings or part numbers, sorted by the user’s own criteria. An AI can tell someone “model A suits a small household.” It can’t work through a three-hundred-row cross-comparison, because those three hundred rows sit in your database, not in its context.

Third: live data — stock levels, remaining places, today’s price, available appointment slots. This changes fast, whatever the AI holds is usually stale, and the models know it, which is why they tend to tell people to check the official site.

The last two are the hardest to displace and the ones people least often realise they already have: the transaction itself (ordering, booking, registering, paying), and anything behind a login — customer records, order history, progress tracking. Crawlers can’t reach that layer at all.

What these share isn’t that they need input. It’s that the answer exists only at your end. The question is whether there’s a place on your page where someone has to hand you their conditions before they can get a result that belongs to them alone.

Which one to build first, by site type

The economics differ sharply by type of business.

Ecommerce: the transaction is already on your site, so you have a third layer. What’s usually missing is filtering — pushing product comparison to a level of granularity an AI can’t stand in for.

Local services (pawnbrokers, clinics, trades, restaurants): start with a calculator that runs on your own pricing bands. Almost every question in these businesses is “what would this cost in my situation” or “can you handle this,” and the answer turns on the customer’s circumstances and your charging rules — and only you hold the second half. One page that produces a number beats ten explainer articles.

B2B and SaaS: start with an assessment tool, but pick one that runs on your own benchmarks — a compatibility check, or a health check that scores their current setup against the sample you hold. A generic ROI calculator carries more risk; that formula is a search away. These double as a lead source, and a high-intent one.

Media and content sites: the hardest case, because content is the product. The workable direction is a queryable dataset — turning what you’ve accumulated into something people can search, compare and filter, rather than another article.

Priorities by site type are broken down further in a separate piece; this one only covers the conversion-layer square.

Don’t start by building a product

The usual failure is scoping it too large: someone hears “build a tool,” and it becomes a spec, an engineering team and a three-month timeline. Halfway through it turns out nobody is using it, and the whole thing stops.

The practical move is to build something that answers exactly one question.

Pick the question your support or sales team fields most often where the answer depends on the customer’s circumstances — the top three usually account for most of the volume. Make it one page: two or three input fields, one number or one clear recommendation out. No accounts, no saved state, no admin panel. A single page of front-end arithmetic.

One trap to avoid: if that page computes something whose formula is public (unit conversion, a standard interest rate), what you have built may only be training material for the AI. Build the kind nobody else can compute — put your bands, your lead times, your surcharge rules inside it, and it stops being portable.

That takes a day or two, and when it’s done you find out immediately whether anyone uses it. If they do, build on it. If they don’t, you’re out two days and you’ve learned that wasn’t where customers were stuck.

One ordering rule: get cited first, then build the thing that catches people. The tool is the bottom of the funnel. If nothing is flowing down from above, the quality of the tool is irrelevant.

The hard part isn’t building it, it’s knowing which one to build

A single calculator page isn’t technically demanding. Everything in this article is within reach.

The awkward step comes earlier. How do you know which question to pick? Going on impressions usually picks wrong — your sales team remembers what they were asked recently, not what they’re asked most. And the questions an AI is now answering on your behalf are ones you have no visibility into at all, because those people never reached your site.

Choosing correctly means first knowing which questions the AI engines are using your content to answer, which ones they answer incompletely, and which ones leave the user still needing a number of their own. That list isn’t in your analytics. You have to ask for it, round after round.

If you want to know which layer your site is missing, we can take a look first.

Further reading