How You Are Seen When the Answer Is Written by a Machine

We wanted to know how far it has gone in Sweden, so we measured instead of guessing. 9,580 Swedish keywords in our industry, with fresh data from September 2026. What you read below is based on that measurement, on Search Console data from our own website, and on studies that have actually measured the effect, not on the estimates circulating in the industry.

Short version: AI answers have taken the informational questions but not the business searches. What determines if you are cited is not your domain value, but whether your content is structured so that an answer can be extracted from it.

AI search is the collective term for searches where the answer is formulated by a language model instead of being presented as a list of links.

To show what it means in practice, we asked the same question: "What is technical SEO" — in several places on the same day, September 23, 2026. The images below are unchanged screenshots from each service.

The box that sits above the regular hits and summarizes the answer with some sources linked. It is available in Sweden and in Swedish, and according to Google cannot be turned off; the closest is to choose the Web filter after the search is done.

Note where the first regular hit lands. It is below the entire AI box, after a button called "Show more." The page ranks as well as before. It just doesn't appear until you scroll.

The sources in the box to the right are mainly two Swedish agencies, and both have written exactly the type of overview we describe below. Clicking on show more reveals a total of six different sources.

A separate search mode, accessible as the first tab in the menu bar on Google. Here, there are no regular hits at all; the entire results page is a conversation, with a box at the bottom where you ask the follow-up question.

The answer is longer and more elaborated than the AI overview, with its own comparison table. The sources have moved to a panel on the right and include a couple of additional sources, a total of ten different ones.

It is the same sites as in the AI overview that mainly appear. The one who is cited is simply cited in both places.

ChatGPT, Perplexity, Claude, Copilot, and Gemini, where the user asks the question directly to the model and never passes a results page.

Here, they differ in a way that determines whether you can be seen at all.

Perplexity provided ten sources and placed a label directly in the body text, plus a panel with all of them.

ChatGPT gave a complete answer with twelve points, a calculation example, and an offer of a checklist — without citing a single source. No site was mentioned, none linked.

Gemini answered in a structured way with its own headings, also without sources in the answer.

That doesn't mean they never cite sources. Both can search the web, and both do so sometimes, depending on how the question is asked, which model is used, and whether the search function is enabled. This particular question, asked in this way, was answered from the model's own memory.

And that's the point. You can't assume a link will follow. In Google and Perplexity, citation is part of the interface. In a chat, it is possible but not guaranteed; your work may have shaped the answer without your name appearing.

Several different interfaces, the same question, and not a single answer that starts with a list of links.

They behave differently, but the problem they create is the same: your content can be read, summarized, and used without anyone visiting your website.

We measured 9,580 Swedish keywords related to digital services — SEO, accessibility, conversion, CMS, e-commerce, design, and development. For each word, we checked if Google shows an AI answer.

AI answers appear on 11.8 percent of the words. Weighted against search volume, it becomes 11.1 percent, roughly one in eight searches.

That figure is on the low end of what is measured internationally, and that's expected: Swedish is a smaller language, and Google rolls out more slowly here. But the average hides the interesting part.

Breaking it down by the type of question asked, it looks completely different:

• Informational questions — what, how, why, when, which: 36 percent
• Our industry's core words — seo, wcag, conversion, cms: 7 percent
• Purchase-related searches — agency, consultant, hire, quote, price: 3 percent
• Local searches — stockholm, gothenburg, near me: 2 percent

That's the whole picture in four lines.AI answers have taken the questions, not the businesses.

If someone searches "what is technical seo" they encounter a machine-written answer. If someone searches "technical seo agency stockholm" they almost never do.

It depends entirely on what kind of searches bring you visitors today.

Ahrefs measured in December 2025 that a first place loses about 58 percent of its clicks when an AI answer is above. In December 2023, the corresponding loss was 34.5 percent. The effect does not diminish over time; it worsens as AI answers appear more often and users get used to not clicking.

France provided the clearest proof, as AI answers had a start date there: July 22, 2026. Ahrefs followed 963 domains in Search Console, 28 days before and nine days after. The most exposed domains lost 23.1 percent of their click-through rate in nine days. The barely exposed ones did not move at all.

The most important thing in that study is not the percentage but the correlation:every percentage point of AI exposure cost about one percentage point of click-through rate.

And a detail that's easy to miss: the impressions did not fall. The pages ranked as well as before. What disappeared were the clicks to the sites.

If you stop following click-through rate and only look at impressions, everything looks normal while traffic drains away.

Here it gets interesting, and here we did a measurement we haven't seen anyone else do on Swedish search results.

We took ten central keywords in our industry, extracted all the sources that the AI answer cites, and compared them with those that actually rank in the top ten organically.

Of 49 cited sources, 27 were in the top ten. So 45 percent of what AI cites is not visible in the top ten for the same search.

An example that stands out: for "customer journey," the AI answer cited seven sources.Only one of them was in the top ten.

The conclusion is uncomfortable for anyone who has bought SEO as placement work, and encouraging for everyone else:being seen in the AI answer is not the same as ranking high. It is a partly separate game, with partly separate rules.

Two additional patterns from the same measurement:

On the word "wcag" and on "accessibility directive" the AI answer cited no external source. Google answered without pointing further. It's the hardest variant of zero click.

And on 96 percent of the keywords showing an AI answer, "People also ask" is also displayed. The two are linked, which is a direct clue about what should be in the text.

We have worked on this on our own website and with clients since AI answers started appearing in Swedish. Five things recur.

A language model cites a paragraph, not a page. If the answer to the question is spread over four sections, there is nothing to cite.

Write the complete answer in one paragraph, early, formulated so that it stands independently. Then the deepening. It's the same discipline as a good press release: the truth first, the background later.

That doesn't mean the text should become a dictionary. An answer with authority is still an answer.

The models match the question against the title. A title that reads "What does an accessibility review cost?" has a chance that "The Investment" never gets.

That 96 percent of the keywords with AI answers also show "People also ask" is no coincidence. Google has already told which questions count in the subject; that list is free title material.

Numbers, dates, prices, and requirements should stand as statements that can be verified, not embedded in reasoning. "The requirements are based on EN 301 549, which in practice means WCAG 2.1 level AA" is citable. "The regulation rests on a European standard that in turn relies on established guidelines" is not.

A section with straight questions and short answers at the end of the article is the format models pick up most easily. But it's not enough to just write it; it must be marked up in the code so that the machine understands that it is questions and answers.

This is the most common deficiency we see. The text exists, the markup is missing, and the page gets nothing for the work.

Named author with title, linked to a page about the person. Sources to the primary source — the authority, the standard, the legal text, and not to someone else's summary.

These are the same signals that have always built trust when it comes to SEO. The difference is that they are now also read by a machine.

1. Measure your exposure. Identify which of your most important keywords show AI answers today. In, for example, Ahrefs Site Explorer, it is available as a filter on SERP features. The proportion is your risk profile.
2. Separate the questions from the business. Keywords that answer a question will lose clicks. Keywords where someone wants to hire, buy, or book hold. Sort your words into two piles and accept that the first pile should be measured on something other than traffic.
3. Set a baseline before you change anything. Without a measurement point before, the effect cannot be shown afterward. This applies to both click-through rate per keyword and how often your brand is mentioned in AI services.
4. Rebuild your most important pages according to the five points above. Start with the pages that already rank but lose clicks; the work is already done there.
5. Change the yardstick. Ranking and sessions are no longer enough. Add the proportion of searches where you are mentioned and the number of citations. Expect the traffic figure to fall while the business remains intact.

One thing that is rarely mentioned: what makes a page readable for a screen reader also makes it readable for a language model.

Sensible heading hierarchy, descriptive alt texts, semantic code, content that exists in HTML instead of being rendered with JavaScript. Same work, two reasons.

We work daily with accessibility and WCAG, and that overlap is one of the more concrete arguments for doing the work now.

"AI search kills SEO." No. Our figures show three percent AI answers on purchase-related searches. The one selling something still has their channel. What changes is that informational content must be measured on something other than clicks.

"You can optimize your way into the AI answer with a trick." There is no trick. There is structure, facts, and credibility — the same thing as always, but with lower tolerance for sloppiness.

"GEO is something completely different from SEO." Mostly not. Technical foundation, content quality, and authority still apply. What is genuinely new is that the answer should be extractable, and the yardstick moves from clicks to citations.

"If you have high domain strength, you'll be fine." Our measurement disagrees. 45 percent of what AI cited was outside the top ten. Structure beats size more often than you think.

Yes. AI overviews are available in Sweden and in Swedish. According to Google, the function cannot be turned off; the closest is to choose the Web filter after the search is done, which shows only text links.

In our measurement of 9,580 Swedish keywords in digital services, AI answers appeared on 11.8 percent. The proportion was 36 percent for informational questions and 3 percent for purchase-related searches.

Ahrefs measurement from December 2025 shows that a first place loses about 58 percent of clicks when an AI answer is shown. In the French study from July 2026, the most exposed domains lost 23.1 percent of their click-through rate in nine days.

GEO, generative engine optimization, is the work of appearing in AI-generated answers. It is based on the same foundation as SEO — technology, content, and authority. But it adds that the answer should be extractable verbatim and that the result is measured in citations rather than clicks.

No. In our review of ten Swedish keywords, 45 percent of the cited sources were outside the top ten organically.

It requires a tool that asks questions to the models and records which brands and sources are mentioned. Ahrefs Brand Radar is one; there are several. The important thing is to set a baseline before the change work begins.

Pages that answer a question that fits in a paragraph. A question that requires a tool, an account, a price, a purchase, or a human assessment does significantly better.

No, if it is not marked up in the code. The text must be supplemented with markup so that search engines and AI services understand that it is questions and answers.

Unsure of how exposed you are? We measure your share of AI answers and what it means for your traffic.

Read more about our work with search engine optimization · Read about our AI solutions

• AI Overviews on Google Search — Google, accessibility and function. Checked September 23, 2026.
• Update: AI Overviews Reduce Clicks by 58% — Ahrefs, December 2025.
• AI Overviews Cut CTR by 23.1% in France — Ahrefs, September 2026. 963 domains in Search Console.
• Own measurement: 9,580 Swedish keywords in digital services, Ahrefs Keywords Explorer, September 2026.
• Own measurement: cited sources in AI answers for ten Swedish industry keywords, compared to organic top ten, September 2026.