How AI Is Changing SEO in 2027

Every year someone announces that SEO is dead, and every year it turns out to have changed shape instead. 2026 was different only in that the change was large enough to show up in the numbers. Going into 2027 the useful question is not whether search still works. It is which parts of it still pay, and what the work looks like now. This is an attempt to answer that from the data rather than from vibes.

What the data actually says

The most-cited evidence is a Pew Research Center study of 68,879 real Google searches from 900 US adults who shared their browsing through a browser extension. When an AI summary appeared, users clicked a result 8 percent of the time. When one did not, they clicked 15 percent of the time. They ended the session entirely 26 percent of the time with a summary present, against 16 percent without. Roughly 1 percent of visits involving an AI Overview included a click on one of the sources it cited.

Google disputes it, and the objection is not unreasonable: the company calls the query set unrepresentative and says it has not seen significant drops in aggregate web traffic. Both things can be true. Aggregate clicks across the whole web can hold up while the click rate on any individual informational query falls, because the number of searches keeps rising. What matters to you is not the aggregate. It is that a ranking which used to deliver a hundred visits now delivers fewer, and that the gap is widest on exactly the informational queries most blogs are written to win.

Nobody left Google. They added a second habit.

The other half of the picture comes from traffic panel data. Traditional search still draws around 3.3 billion monthly unique visitors worldwide. Every AI chatbot combined draws about 655 million — roughly a fifth. And 95 percent of ChatGPT users were also using Google in the same period. People did not switch. They added a tool.

That is the single most important framing for 2027 planning. If you treat AI search as a replacement channel you will over-invest in it and neglect the thing still sending most of your traffic. If you treat it as a second surface where the same groundwork pays twice, you will spend your time correctly.

Assistants still barely link out

In May 2026, around 6.8 percent of ChatGPT answers contained an external link. That is a fivefold rise from roughly 1 percent a year earlier, and it is still small. Ninety-three of every hundred answers send traffic nowhere. Meanwhile AI platform visits grew sharply year on year while referral traffic out of those platforms stayed broadly flat. More usage, not proportionally more clicks.

So being cited is mostly not a traffic play. It is a brand play that converts later, through a search for your name or a direct visit. Which leads to the most useful statistic in all of this.

Deep pages earn the citation. The homepage gets the visit.

Around 65 percent of URLs cited by assistants sit two or three folders deep — guides, comparisons, specific answers. But close to 59 percent of the referral traffic that does arrive lands on homepages. The person reads the cited guide inside the assistant, does not click, remembers the brand, and arrives later through the front door.

That breaks the normal attribution model completely. Your deep content is doing the acquisition work and your homepage is getting the credit. If you judge your guides on their own landing-page traffic you will conclude they failed and delete them, which is precisely the wrong move. Judge them on citations earned and on brand search volume over the following quarter.

Fewer visitors, better visitors

When AI referrals do arrive they behave well. Reported figures put them around a 7 percent conversion rate on transactional sites against roughly 5 percent for conventional search referrals, with longer sessions and more pages per visit. That makes sense — the assistant has already done the filtering, the comparison and some of the persuasion before the person clicks.

The practical consequence is that sessions became a bad headline metric some time in 2026. Revenue per session is the number that still means something. A site whose traffic fell 30 percent while revenue held flat did not have a bad year.

What actually changes in 2027

Five shifts are already underway and will be ordinary by the end of next year.

1. Measurement moves from rankings to share of answer

Position tracking keeps working for conventional results and tells you less every quarter. The replacement is unglamorous: a fixed panel of the questions you want to own, run monthly through each assistant in a clean session, recording whether you were named, who was named instead, and what the answer said about your category. It is manual, it is boring, and it is the only measurement in this area that is not an estimate. Build the panel now so that by mid-2027 you have a year of trend.

2. Pages get written for two readers at once

The structure that wins citations is a specific question as a heading, followed immediately by a complete two-to-four-sentence answer, followed by the depth for the human who wants it. That is not a trick, and it is the same shape we set out in how to get cited by ChatGPT and Perplexity. It is just the page a model can quote without reconstructing your argument, and it happens to be a page people prefer too. Expect the long throat-clearing introduction to disappear from competently run sites during 2027.

3. Brand search becomes the conversion funnel

If assistants describe you to people who never click, the measurable consequence shows up as branded search and direct traffic. Those become leading indicators rather than vanity metrics. Watch branded impressions in Search Console as a proxy for how often you are being mentioned somewhere you cannot see.

4. Agents start transacting, not just answering

The move from assistants that recommend to agents that compare, fill carts and complete checkout is the genuinely new thing for 2027. It is early and it is overhyped, but the direction is clear and the preparation is cheap, because an agent reads the same structured product data a search engine does. Accurate price and availability, clean feeds, stated shipping and returns terms, no information trapped in images or JavaScript. Retailers who already did their schema properly will find they are mostly ready, and what that means for adult stores specifically is set out in agentic commerce for adult retail.

5. The crawl becomes contested

Through 2026 a growing number of publishers began blocking or charging AI crawlers, and the licensing fights moved into court. Expect more fragmentation in 2027 — some content reachable to some models and not others. For most businesses the decision is simple and in one direction: if you sell something, you want to be readable. Blocking AI crawlers is a strategy for publishers whose product is the content itself. If your content is marketing, blocking is self-harm.

What is not worth your time

llms.txt remains unsupported. Google has described it as speculative, no major assistant has committed to reading it, and adoption outside SEO circles is negligible. It costs ten minutes to add, so add it if it makes you feel better, but do not let anyone sell it to you as a service.

Subscriptions promising AI visibility are mostly reselling what you can measure yourself with a spreadsheet and an hour a month. Publishing high volumes of AI-written content is the same losing move it was before, just cheaper to execute and therefore more tempting. And nobody needs a separate GEO strategy that contradicts their SEO strategy — if the two are different documents, one of them is wrong.

What this means in restricted industries

Everything above applies, plus one layer that does not exist for mainstream businesses: the refusal. Assistants apply content policy before they apply retrieval. Explicitly adult queries get declined, deflected, or answered without named sources, and no amount of technical work changes that. If your entire content plan targets explicit intent, AI search will never carry it.

The opening is the informational layer around the industry — safety, materials, legality by jurisdiction, platform comparisons, how the business side works. Those questions get answered and cited, the competition for them is thin because most operators in adult, casino and cannabis have not started, and they sit directly upstream of commercial intent. That window is wider in 2026 than it will be in 2028.

The second restricted-industry issue is access. Age gates, interstitials and aggressive bot blocking stop crawlers and agents at the door. A gate that serves nothing but the gate gets you indexed as nothing but a gate. Keep the informational content on clean, reachable pages and keep the verification where it legally has to be, not in front of everything.

The uncomfortable summary

Rankings are worth less per position than they were, and that is not going to reverse. In exchange, the traffic that still arrives is better qualified, the number of businesses doing the work properly is small, and the groundwork that wins citations is the same groundwork that has always won rankings — crawlable pages, specific answers, a consistent identity, and other people referring to you.

If you were doing it well in 2024 you are mostly fine. If you were getting by on volume and thin pages, 2027 is the year that stops working entirely. If you would rather hand it over, that is what our AI SEO service covers.

Sources

Click-rate figures come from the Pew Research Center study of 68,879 Google searches, reported with Google’s response in PPC Land. Audience, citation and referral figures come from Similarweb’s AI search data and Search Engine Journal’s analysis of it.