Resources / Statistics
Jul 2026 Last checked
Yes. Around half of visitors reach for search first, and most sites fail the queries they type. Five of these seven come from the same benchmark, so read them as one picture rather than seven independent findings.
Search failure is the normal state, not the exception. If you have never tested your own search against real queries, assume you are in this group.
Put next to the number above, this is the gap: a third of people use the search box for questions, and two thirds of sites cannot answer them.
Search boxes are usually wired to the product catalog only. Policy and support questions return nothing useful. This is the worst-performing of the eight query types tracked.
Search failure does not produce a complaint. It produces an exit, and you never learn which query caused it.
Search is not a minor feature used by a fraction of visitors. It carries about half of all product-finding attempts.
These queries carry the most buying intent on the site. Someone asking whether a part fits their model has already decided to buy something.
Yes. Roughly two thirds of US adults now meet AI summaries in search results at least sometimes.
Close to half the adult population meets this format often enough to have formed expectations about it.
Reading an AI answer instead of a list of links is now ordinary behavior for most adults, not an early-adopter habit.
Do not plan around this as a traffic channel yet. Plan around it as a signal about what visitors have already read before they land.
If your customers skew young, they already expect to type a full question and get a written answer back.
The load is real and poorly served. Self-service fully resolves only 14% of issues, and the most common reason it fails is that people cannot find the content. We could not find independent research on AI ticket deflection rates. What follows measures the underlying problem instead, which is far better evidenced.
Most people who start in self-service finish somewhere else. The distance between starting and finishing is where support cost accumulates.
The content usually exists. The finding is what fails. That is a retrieval problem, not a writing problem.
Half of all support contacts generate more support contacts. That is the volume any deflection strategy is aimed at.
The expectation moved in twelve months. Whatever felt fast last year is now the baseline.
Three quarters of your customers expect an answer at 2am. Most of those answers are already written somewhere on your site.
People use AI answers without trusting them. One in five call them very useful, and nearly half report little or no trust.
Adoption is not enthusiasm. Most people use AI answers without rating them highly.
More than a quarter get no value from the format. Any AI answer you put in front of customers has to clear a bar that general-purpose summaries currently do not.
Dissatisfaction is near universal and has survived several years of AI deployment. Adding AI has not by itself moved this number.
The interaction matters more than the channel it arrived through. An answer that lands in two seconds counts as a positive experience.
Clicks roughly halve. When an AI summary appears, the click rate on a result falls from 15% to 8%. This section uses browsing data rather than a survey. Pew tracked what people did, not what they said they did.
Consumers say it does, though the available evidence comes from companies that sell translation services. Read this section with care. All four numbers come from research commissioned by a company that sells translation services, and the questions are framed to invite agreement. Treat the direction as sound and the magnitudes as generous.
