When the AI answers instead of your website: how to keep getting customers

Someone searches «bathroom renovation in Valencia». Three years ago they saw ten blue links and opened two or three. Today they read a ready-written paragraph with rough prices and three tips, and often close the tab without clicking anything. Your site was there, among the sources. Nobody ever saw it.

That scene is the real shift of the past year, and it is not fixed by publishing more articles.

What is actually changing?

Two things at once. First, search stopped being an index of links and became a place where you read the answer: generated summaries above the results, plus a conversational mode for harder questions. Second, a lot of people no longer start at a search engine at all — they ask an assistant, which answers and cites a few sources.

The combined effect has had a name for years: zero-click search. It used to affect trivial questions — the weather, a phone number, a unit conversion. It now reaches questions with commercial intent, which are precisely yours.

How much traffic is actually lost?

The most citable public measurement is from Pew Research Center, July 2025, based on real browsing behaviour: when the results page included a generated summary, users clicked through on roughly 8 % of visits; without one, around 15 %. Call it half the clicks.

Google, for its part, maintains that the clicks that do come through are higher quality, because the user has filtered first. Both statements can be true at once — and in practice that is what you see: fewer visits, further along in the decision.

What you should not do is adopt either number as your own. The symptom takes ten minutes to check in Search Console: if impressions hold steady while clicks fall month after month, it is already happening to you. If branded traffic is also rising while generic traffic falls, it is happening all the more.

The uncomfortable arithmetic: fewer visits, and dearer

Here is the sum almost nobody says out loud. If your site took 1,000 organic visits a month and converted at 2 %, that was 20 enquiries. If the same content drops to 600 visits, you now need 3.3 % conversion just to stand still. Nobody is going to hand you that extra point and a half — you have to win it on the page.

And if you make up the shortfall with ads, cost per visit rises, because you are bidding for the same inventory as everyone else. Which is why, in 2026, the cheap part is no longer getting people in. It is not wasting the ones who arrive. The form-versus-quiz conversion data we collected in quiz vs form: what the data says is worth considerably more today than two years ago, simply because the numerator has shrunk.

How to get into the answer, not just into the index

An assistant cites what it can extract and verify. Obvious enough — and it disqualifies half of what sits on service-business websites: prose without data, promises without figures, prices that only ever appear as «on request».

• Put the facts in text: price range, service area, typical lead times, what is included and what is not. If the price lives only in a PDF or behind a button, for the model it does not exist.
• Keep a visible FAQ on the page, written the way customers actually ask, answered in two sentences. The same information in structured data helps, but what gets quoted is the text.
• Write headings as questions and answer immediately underneath. A self-contained paragraph can be quoted; one that depends on the previous paragraph cannot.
• Look after your Google Business Profile and real reviews: for local searches it remains the heaviest source, with or without AI.
• Date it and keep it current. Between two equivalent sources, the one with a recent date and concrete numbers wins almost every time.

On llms.txt-style files — a proposal for handing models a plain-text map of your site — the honest answer is this: there is no public evidence that engines use it to decide whom to cite. It costs half an hour, it does no harm, and for us it was useful mainly because building it exposed inconsistencies in our own numbers. But it does not replace the text on the page, and anyone selling it to you as a ranking trick is selling smoke.

What does not work (and now costs more)

• Publishing fifty generated articles with no data of your own: that is exactly the material the model can already write by itself.
• Hiding the price to «force» a phone call. No figure, no citation — and the customer asks someone else.
• Inventing ratings or stars in your structured data: besides breaking the structured-data guidelines, a generative engine will repeat that lie with your name attached.
• Optimising keywords for positions that no longer receive clicks, instead of answering the question behind them.

The good news: the visitor who does arrive is closer to buying

If the assistant already explained what a turnkey bathroom refit involves and roughly what it costs, whoever clicks your link is not there to learn. They are there to decide. That shift makes two old things decisive.

First, do not meet them with an eight-field form. Every extra field at that moment is an exit — we go through it in the leads you lose before the submit button. Second, answer fast: that same visitor is usually already talking to two other suppliers, so speed decides more than your pitch, as we saw in the first five minutes.

What to do this week

• Open Search Console and compare impressions against clicks for the last six months. Write down the trend — that is your baseline.
• Take your three highest-impression pages and add an indicative price, service area and lead time, in text.
• Add five genuine frequently asked questions, each answered in two sentences.
• Replace the form on your main page with two or three qualifying questions and move contact details to the end.
• Measure where people drop out and what share reach the last step.
• Set an internal response promise — fifteen minutes in working hours — and test it yourself with a dummy lead.

None of those six is a new technique. What is new is that they now pay for themselves on half the traffic.

How we do it at Quizanta

We apply to our own site exactly what you just read: prices and plan limits are in text, there is a visible FAQ with matching structured data, and we publish a plain-text map for models. The product handles the other half of the problem: every visit becomes a short quiz that qualifies, scores and delivers the lead to WhatsApp the moment it is completed. Less traffic does not have to mean fewer customers — it means you can no longer afford to lose the ones who arrive.