SEO and AI meet at three different points: artificial intelligence as a tool for doing SEO work, optimisation to appear inside the answers generated by ChatGPT or Gemini, and the effect Google's AI Overviews have on the traffic a site already gets. They are three separate problems, with three separate metrics. Mixing them up is why many companies spend on the wrong thing.
Below is what the measured data says on each of the three fronts, and what still works exactly as before.
Behind "SEO and AI" there are three different jobs
When a company asks to "do SEO with AI", the same sentence nearly always holds three incompatible requests. Separating them before signing a proposal saves you from paying for one job and expecting the results of another.
| The job | What gets optimised | How it is measured | Who benefits straight away |
|---|---|---|---|
| AI as a tool | The internal process: keyword research, clusters, drafts, log analysis | Hours saved for the same published output | Teams producing a lot of content with few people |
| AEO / GEO | The content, so that assistants cite it | Mentions and citations in answers, referral traffic from chat | Businesses selling products or services people ask for recommendations on |
| Defending against AI Overviews | The mix of queries you target, and the page that receives them | Clicks for the same impressions in Search Console | Sites that already have informational traffic to protect |
The distinction matters because the three jobs run on different timescales. The first pays back in weeks, the second in months, the third is ongoing maintenance that never ends. We covered the second at length in the guide to generative engine optimisation and in the piece on what answer engine optimisation is, where the differences between AEO and SEO are explained one by one.
How much AI has eroded clicks: the verified numbers
Very different figures circulate on the fall in traffic, often without a source. These are the ones that come from published studies, with the sample stated.
| Study | Sample | Result |
|---|---|---|
| Pew Research Center, July 2025 | 900 US adults, real browsing in March 2025 | With an AI summary on the page, people click a result on 8% of visits, against 15% without. Only 1% click a link inside the summary. |
| Pew Research Center, same study | As above | 26% end the session after a page with an AI summary, against 16% without. |
| Ahrefs, 2025 | 300,000 keywords, half with an AI Overview and half without, March 2024 compared with March 2025 | Position one CTR falls by 34.5% when an AI Overview appears. |
| Visibility Labs for Search Engine Land, 2025 | 94 e-commerce sites, 9.46 million non-brand organic sessions against 135,000 sessions from ChatGPT | ChatGPT traffic converts at 1.81% against 1.39% for non-brand organic, 31% higher. But it accounts for 1.48% of total organic revenue. |
The rows to read together are the last one and the first ones. Traffic from assistants converts better, and that figure is quoted everywhere. What gets quoted much less is the second half of the same sentence: in the Visibility Labs sample it accounts for 1.48% of organic revenue, rising to 2.2% in the second half of 2025. It is a share that is growing fast and that does not yet pay the rent.
For a B2B company the practical consequence is that investment has to be calibrated on two horizons at once, not moved wholesale. Anyone who has stopped working on organic rankings to chase citations in assistants has given up 98% of the channel for 2%.
Using AI to do SEO: where it holds up and where it does not
In the UK adoption has tripled in under three years. According to the ONS article Artificial intelligence in UK businesses: 2023 to 2026, based on the Business Insights and Conditions Survey, about 35% of businesses with 10 or more employees used at least one AI technology in June 2026, against about 12% in September 2023. The figure is around 28% for businesses with 10 to 49 employees and 49% for those with 250 or more. The most widely used technologies are large language models (18%) and visual content creation (16%), both close to the day-to-day work of marketing.
In daily work the dividing line is clear. AI is reliable on anything that transforms data you already have: grouping thousands of Search Console queries into semantic clusters, rewriting 400 title tags to a pattern, reading a log file, comparing two exports. It becomes unreliable as soon as you ask it to produce the fact rather than process it.
The sensitive point is numbers. A model that writes "67% of companies say" without a source in its input is completing a sentence, not reporting a statistic. Every figure has to be traced back to the original study, and if the study does not exist the sentence comes out. The same applies to sources quoted by others: if a blog post reports a Semrush figure, link to Semrush, not to the blog post.
There is less mystery about Google's position than people suggest. The Search Central guidelines reward useful content regardless of how it was produced, and penalise large-scale production aimed at manipulating rankings. The deciding factor is not the tool, it is whether the piece adds something. We have also written about this in the pieces on AI detectors and E-E-A-T.
Getting found inside the answers
Assistants build their answers by drawing on sources they can read and trust. The levers you can control are few and concrete.
- A short, self-contained answer in the first paragraphs, one that still works when lifted out of the page.
- Data in HTML, not inside images. To a language model, a table in a screenshot is a smudge of pixels.
- Correct structured data, particularly Article and FAQPage, which make the structure explicit instead of leaving it to interpretation.
- A real author, with a bio explaining why they are qualified to write on the subject. Assistants favour attributed content.
- Brand mentions outside your own site, which feed the models' trust far more than backlinks alone.
For the operational detail on each assistant we have published separate guides: how to do SEO for ChatGPT, SEO for Perplexity, SEO for Gemini and what changes with Google AI Mode. We wrote a separate guide on the llms.txt file, with the caveat that no major provider has yet declared support for it.
What has not changed: the Macropix case
Macropix makes custom LED screens in Milan. When the project started in 2020, the site did not appear for any of the category keywords in its sector on Google Italy. For "monitor pubblicitario" (advertising monitor) it was in position 88, on page nine.
By 2025 that same keyword was in second position. "Ledwall", worth 4,400 searches a month, went from absent to first. In July 2026 the domain held 25.55% share of voice on the ledwall cluster, ahead of Amazon with 18.08%. On the paid side, between July 2025 and June 2026, the account produced 374 conversions from €16,805 invested, at €44.93 per conversion, with conversions up 84.2% on the previous twelve months.
None of these results comes from anything to do with AI. They come from a site architecture built around production capabilities, low-volume technical keywords and content that answers purchasing department questions. The full method is in the piece on SEO for manufacturing, and the full project in the Macropix case study.
The reason this case belongs in an article about AI is what happens to technical, niche queries. On Google Italy, where Macropix competes, AI Overviews rarely appear on industrial supplier searches: in a check we ran in September 2026, none of the five supplier and price searches had one. On Google UK the picture is different. In our check of ten UK B2B queries, nine had an AI Overview, supplier searches included, although on several of them it sat below the local pack or at the bottom of the page. The pages that work there are the same ones that worked for Macropix: specific, technical, written by people who do the job, which are also the pages the summaries cite. We covered the numbers in the article on the impact of AI Overviews on SEO, and the wider question in will AI replace SEO?.
How to measure it now
Average position on its own has become a dangerous metric: you can move up and lose clicks, because above the first result there is a block that answers the question. The three readings to put on the dashboard:
- Clicks for the same impressions. Stable impressions with falling clicks on an informational page is the typical pattern of an AI Overview taking the answer. In Search Console you see it by filtering by page and comparing two like-for-like periods.
- Referral traffic from assistants. In analytics, isolate the sources chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com. The volumes are small and should be read as a monthly trend, not an absolute number.
- Presence in answers. Test by hand ten or fifteen prompts a customer would really use, on the three or four main assistants, and note whether the brand appears and how it is described. Done once a month, it becomes a time series. There are AI visibility tools that automate the monitoring.
The third check is the one almost nobody does, and it is also the only one that tells you whether the AEO and GEO work is producing anything.
Where to start
If you are deciding where to put next quarter's budget, the order that makes most sense for a company starting from a normal position is this.
- Open Search Console and isolate the pages with stable impressions and falling clicks over the last twelve months. That is your real damage, quantified.
- On those pages, move the content to where the AI Overview does not reach: cases with numbers, operating procedures, comparisons between options. Google has already taken the definitions.
- Sort out author, bio and structured data across the whole archive. It is low-cost technical work that serves both classic SEO and citations in assistants.
- Only at this point bring AI into the production process, and only for data transformation tasks.
If you would rather start from an analysis of where you stand, our SEO for AI service exists for exactly this, and traditional SEO remains the piece everything else rests on.
Frequently asked questions
No. Google Search Central guidelines assess how useful content is, not how it was produced. What is penalised is large-scale production with the sole aim of manipulating rankings, whether the author is a person or a model.
It depends on the type of query. The Ahrefs study of 300,000 keywords measures a 34.5% fall in position one CTR when an AI Overview appears, and Pew Research Center finds clicks on results in 8% of visits with an AI summary against 15% without. The erosion is concentrated on generic informational queries.
Not as an alternative. In the sample of 94 e-commerce sites analysed by Visibility Labs, ChatGPT traffic converts 31% better than non-brand organic, but it accounts for 1.48% of total organic revenue. It is a fast-growing share that does not replace the organic channel today.
SEO aims to rank a page in search results. AEO (answer engine optimisation) aims to have it used as the answer by answer engines. GEO (generative engine optimisation) aims to have it cited by generative models such as ChatGPT or Gemini. The techniques overlap a lot, the metrics do not.
According to the ONS, about 35% of UK businesses with 10 or more employees used at least one AI technology in June 2026, against about 12% in September 2023. Among businesses with 250 or more employees the share is 49%. Large language models are the most used technology, at 18%.
One thing the industry rarely says: the most concrete advantage AI brings to SEO right now is not producing content, it is reading the content you already have. An archive of two hundred articles that nobody has reread in years nearly always contains ten pages cannibalising each other and thirty worth updating. Finding them by hand takes days; with a Search Console export and the right model it takes an afternoon. It is the least talked-about job, and on the projects where we have done it, the one that has moved things most.
