Pew found that when an AI summary appears, users click a traditional result on 8% of visits instead of 15%. The response is not to abandon SEO — it is to change what you publish and how you measure it. Here is the practical version.
The question every marketing team is now asking out loud: if Google answers the query on the results page, what is the point of ranking?
It is a fair question and it has a specific answer, which is less dramatic than the headlines and more demanding than business as usual.
What the data actually shows
The most rigorous public study remains Pew Research Center's analysis of 68,879 Google searches from 900 US adults. Its findings are worth stating precisely, because they are routinely exaggerated in both directions.
- Around 18% of the searches studied produced an AI summary.
- When a summary appeared, users clicked a traditional search result on 8% of visits, against 15% when no summary appeared.
- Users clicked a source cited inside the summary on about 1% of visits.
- 26% of visits with a summary ended the browsing session entirely, against 16% without.
Read that carefully. Clicks roughly halve on affected queries — a serious hit, not an extinction event — and the majority of searches in that dataset produced no summary at all. Meanwhile Google's AI Mode passed a billion monthly users by its I/O announcement in May 2026 and is available in more than 200 countries, while Google has said publicly, twice in 2026, that it does not intend to make it the default search experience.
So: a large and growing share of informational queries now resolve without a click, and commercial and navigational intent still largely behaves as it did.
What still earns the click
Look at the queries where a summary either does not appear or does not satisfy, and a pattern emerges.
Commercial intent. Someone comparing vendors, checking pricing, or deciding whether to hire is not finished after a summary. They want evidence, specifics and someone to talk to.
Anything requiring trust. Legal, medical, financial and high-consideration purchases send people looking for the source, not the synopsis.
Original data. A summary can only summarise what exists. If you publish the benchmark, the survey, the teardown or the case study, you are the thing being summarised — and cited.
Tools and interactive content. A calculator, configurator or checker cannot be answered away. It has to be used.
Genuine expertise and perspective. Judgement, trade-offs, opinionated recommendations, first-hand accounts of what happened on a real project. AI systems are good at consensus and poor at conviction.
Freshness. Anything that changes — prices, regulations, platform deadlines, release notes — pulls people to a source that is dated and current.
How to be the source that gets cited
Citation behaves like a related but distinct ranking problem, and the practical levers are unglamorous.
Answer the question early and directly. Put a clear, complete, self-contained answer in the first 100 words under a heading that matches the question. Extraction systems reward passages that stand alone; so do readers.
Structure for retrieval. Descriptive H2s phrased as the questions people ask, short paragraphs, real lists and tables for anything enumerable. A wall of undifferentiated prose is hard to quote from.
Be specific and attributable. Numbers, dates, named methods and named sources. Vague claims are unquotable; a sentence containing a figure and its provenance is exactly what gets lifted into an answer.
Keep entity signals clean. A complete Organization schema, consistent naming, real author identities with credentials, an about page that says who you are. Systems that assemble answers need to know whether you are a credible source, and they read structure to work it out.
Maintain rather than accumulate. Update dates that mean something, deprecated advice removed, superseded pages consolidated. Stale content is not neutral now; it is actively deselected.
Keep earning links and mentions. Every study of AI citation behaviour so far points back to the same underlying signals of authority that conventional SEO has always chased. The system is different; the substrate is not.
How to measure it
This is where most teams are stuck, and it is worth being honest about the tooling.
Search Console's generative-AI reporting arrived in June 2026 and rolled out worldwide at the end of August, but it reports impressions only — no clicks, no queries, no position. It tells you that you are appearing, not what it is worth.
So build the measurement around what you can see.
- Segment your reporting by intent. Split informational from commercial URLs and track them separately. Aggregate organic traffic hides the story completely: informational clicks falling while commercial clicks hold is a healthy site adapting, and it looks like decline on one chart.
- Watch clicks per impression by query type. A falling CTR at stable position is the fingerprint of an answer appearing above you.
- Track branded search. If AI answers are exposing people to your name without a click, branded searches rise. That is the compounding return on this work.
- Sample manually. Check a fixed set of 20 to 30 priority queries monthly, recording whether an AI answer appears and whether you are cited. It is unfashionable, tedious and currently the most reliable data you can get.
- Measure assisted conversions, not last-click. The click you lost was probably the first touch, not the deciding one.
What we would actually do this quarter
Audit by intent. Which of your top 50 organic URLs answer a question in a paragraph, and which help someone decide? The first group is your exposure.
Rewrite the top ten commercial pages to be quotable. Direct answers under question-shaped headings, specifics with sources, a clear next step.
Commission one piece of original data. Your own benchmark, survey or teardown. It is the single highest-leverage thing you can publish in this environment, and it is the sort of asset that keeps returning citations for years.
Shift some budget to channels you own. Email, community and direct relationships. The lesson of the last two years is not that SEO is dead — it is that a business dependent on one intermediary for demand is fragile. This is exactly the case for pairing SEO with email marketing rather than treating them as separate budgets.
Do not abandon the channel. Organic search still delivers the highest-intent unpaid traffic available to most businesses. Halved clicks on informational queries is a reason to change the mix, not to stop.
Questions we get asked
Should I block AI crawlers? Only with a clear commercial reason. Blocking removes you from answers you might otherwise have been cited in, and does nothing to restore the click that the answer replaced. Publishers with licensing leverage are a different case from most businesses.
Does an llms.txt file help? There is no evidence yet that major systems act on it, and it costs almost nothing to publish. Treat it as a low-cost hedge, not a strategy.
Is this the end of blogging? It is the end of blogging that restates what is already known. Original reporting, opinionated analysis and genuinely useful tools are worth more now, not less.
Should we move budget to paid? Paid search is more attractive when organic clicks compress, and rising demand for that inventory is exactly why pay-per-click management has to be measured against qualified pipeline rather than clicks. Move deliberately, not reflexively.
Where AI answers are least likely to intercept you
It helps to know which parts of your keyword set are exposed and which are not. Broadly, an AI answer is least likely to satisfy the user when:
- The answer depends on the user's situation — pricing that varies by scope, eligibility that depends on circumstances, compatibility that depends on what they already own.
- The query is transactional. Someone ready to buy, book or brief wants a supplier, not a summary.
- The information changes faster than a model's training — platform deadlines, regulatory dates, release notes, availability.
- Trust is the actual currency. People want to see the practice, the credentials, the case study and the name of a person.
- The output is a thing to use. Calculators, templates, configurators and checkers cannot be paraphrased into a paragraph.
Audit your top queries against those five conditions. What is left over — the definitional, the encyclopaedic, the how-does-this-work explainer — is the part of your traffic to assume will keep compressing, and the part to stop investing new budget in.
The honest summary: the queries that were easy to win were the ones with the least commercial value, and they are the ones being absorbed. What remains is harder, more expensive, and worth considerably more — which is a description of most of the work that has ever been worth doing in content marketing.
Looking for more? Browse all resources.