AI citations and the future of SEO

When I needed a barber in Gdańsk, I asked Gemini.

I wanted a good place I could actually visit. The recommendation appeared to draw on a combination of reviews, a working website, opening hours, prices and booking information. It brought those details together into a decision I could act on.

Later, the barber told me that other customers had also arrived because an AI assistant recommended the shop. That conversation stayed with me. A business could be discovered, compared and chosen before a customer had opened its website.

It made a change in my own project easier to recognise: people were finding SamfunnPrep through AI, too. I began paying closer attention to what was being indexed, what was being cited, and whether that visibility brought people into the product.

This is my case study of search engine optimization (SEO) and generative engine optimization (GEO): making a project discoverable and useful in AI-generated answers. I use SamfunnPrep’s search and citation reports to examine what changed, what those measurements can tell us, and what I would build next.

The caseA small niche, a measurable footprint

SamfunnPrep focuses on integration and life in Norway. The audience is specific: people trying to understand a country, its institutions and the practical steps of settling there. They may need the same information in very different languages.

Over roughly six months, we produced 799 articles for a project available in 11 languages. The opportunity I see is in covering real questions in enough depth, and making those answers accessible to the people asking them.

To examine the result, I exported the Google and Bing reports on 1 November 2026. Google covers 31 July–30 October; Bing covers 1 August–30 October. Here is what the daily rows actually add up to.

The clearest shift was in Google’s AI impressions

During 1–20 September, the generative AI report averaged 65.7 impressions a day. During 1–30 October, the average was 2,285.0 a day. The highest day was 22 October, with 3,866 impressions.

Figure 01

Google’s AI visibility changed scale

Impressions per day · 31 Jul–30 Oct 2026

3,866peak · 22 October

Daily impressions7-day average

Of the 76,883 impressions in the full export, 68,551 arrived during the first 30 days of October. About 89% of the period’s recorded AI visibility was concentrated in that month. The shift was substantial, even though daily impressions fell back from their peak before the reporting window ended.

Google defines these impressions as occasions when links to a site are shown in its supported generative AI Search features, including AI Overviews and AI Mode. The export does not tell me how many people registered, whether a response recommended the product, or what caused the increase. [1]

Search traffic grew alongside it

The Google web-search export records 6,466 clicks; the Bing Search Performance export records 2,629. October’s increase in Google clicks is visible even though the export stops on the 30th. Bing’s growth is steadier.

Figure 02

More search clicks reached the site

Monthly clicks · August–October 2026

Google web searchBing Search Performance

These trends happened together. They do not establish that more articles, translation, or any individual technical change caused the growth. This is an observational case study of one site, without a control group. Changes in demand, the search products and their reporting could also affect the result.

At the time of writing, my sitewide indexing count was roughly 16,700 pages, including more than the article collection. An indexed URL is an opportunity to be discovered. The useful questions are which pages answer a need, which receive visibility, and what readers do next.

Citations reveal another kind of visibility

The Microsoft AI export records 97,836 citations. Its daily “Cited Pages” column averages 212.2: about 212 distinct pages cited on an average day, rather than 212 unique pages across the entire period.

Microsoft’s report covers Copilot, AI-generated summaries in Bing and selected partner integrations. A citation means that content was displayed as a source. It does not measure authority, ranking or a visit to the website, and it cannot be presented as a count of ChatGPT recommendations. [2]

Figure 03

A recurring presence in AI answers

Citations per day · 1 Aug–30 Oct 2026

97,836citations in the export

Daily citations7-day average

For me, this is evidence that a focused resource can become a recurring source in AI answers. The practical value still depends on what those appearances lead to.

How discovery worksBeing found is different from being learned

My original intuition was that publishing useful information now might shape what future models know about a project. That remains an interesting possibility. But the search results above describe visibility today, and that requires separating three processes.

01 Training

Information may be used when a provider builds or updates a model. A website owner generally cannot verify inclusion in training from a citation report.

02 Discovery and retrieval

A system finds relevant pages and brings information into the context of a particular answer. Fresh information can be retrieved without waiting for a new foundation model.

03 Citation and recommendation

An answer may link to a source. Recommending its product is a further decision. Being quoted for one fact does not mean the assistant endorses the business.

Google describes grounding responses in pages retrieved from its Search index, sometimes using several related searches to answer a question. That means established SEO work still matters: the system needs to find and understand the page before it can use it. [3]

OpenAI also separates search access from training access. OAI-SearchBot is used for search discovery; GPTBot crawls content that may be used for training. Their controls are independent. Allowing one is not the same decision as allowing the other. [4]

User-agent: OAI-SearchBot
Allow: /

A chatbot retrieving a page during a conversation does not, by itself, show that its underlying model learned the page permanently. Likewise, none of my exports establishes that the internet is nearing the end of being “learned”, or that we are at a final peak of model training. I would treat that as speculation, not the premise of a content strategy.

The part we can influence now is more concrete: publish information worth retrieving, let the appropriate systems access it, and keep it accurate when someone needs it.

Visibility should come with a choice

There is another side to this opportunity: the publisher’s choice. I may want an assistant to find my work, cite it and bring readers to my product, while choosing to withhold that work from model training. A strategy for AI discovery should make room for that decision.

On 30 October 2026, Google and OpenAI announced plans for a shared robots.txt signal for AI training preferences. Website owners could communicate one preference to multiple participating providers, reducing the need for separate instructions. As of this article’s publication on 1 November, implementation is still ahead: OpenAI aims to introduce support in the coming months, and Google is targeting 2027. [5] [6]

The proposed mechanism remains a draft, not an adopted IETF standard. It expresses a usage preference that participating providers agree to honour; it does not physically prevent access to a public page. [7] Existing provider-specific controls still matter, and their effects differ. Google plans to treat the shared opt-out as disallowing its Google-Extended control when no explicit rule for that control exists. This also affects grounding in Gemini Apps. [5]

For me, this adds another responsibility to publishing: decide which uses of the content I want to permit. Being discoverable, contributing to training and appearing in an AI answer deserve separate consideration, with attention to how each provider handles those choices.

What to buildA content system with a maintenance habit

I still like the idea of a content factory, provided its output is useful knowledge and its responsibilities extend beyond publication. Producing a draft is only one step. Research, translation, checking, revision and retirement of obsolete information all belong in the same system.

For a project about life in Norway, a page can become less useful when a requirement changes or an institution updates its guidance. Publishing it in several languages makes keeping those versions aligned a real part of the work.

My approach going forward would centre on five practices.

1. Start with a real question.

   Write for the person trying to complete a task. The Google AI page export includes articles about military service, permanent residence, UDI waiting times, alcohol rules and unemployment benefits. These are specific concerns with practical consequences. They are a better starting point than a quota of pages.

2. Make the answer checkable.

   Identify the relevant country, audience and date. Link to the primary source of a rule. Distinguish an official requirement from an explanation or personal experience. Where the answer depends on circumstances, explain those circumstances instead of smoothing them away.

3. Treat each language as a responsibility.

   A translated page should preserve the meaning of the source and use terminology a reader understands. I would keep a shared record of the facts behind each article so a change can be reviewed across its language versions.

4. Keep discovery technically straightforward.

   Give important pages stable URLs, useful internal links and accessible text. Check indexing controls and duplicate URLs. Review the Search generative AI inclusion setting in Google Search Console. Keep sitemaps current and use supported change-notification mechanisms where appropriate. Google says normal SEO foundations still apply, and does not require special AI markup or an llms.txt file for its Search features. [3]

5. Measure outcomes, then revise.

   Track search clicks, AI visibility and registrations as separate measures. Preserve attribution where available. Compare consistent periods and keep a record of substantial content changes. A useful next experiment would compare updated articles with similar unchanged ones, while acknowledging that search demand and platform changes remain outside my control.

AI can help research a structure, prepare a draft or translate an explanation. The publishing decision still needs a standard for accuracy and added value. Google’s guidance specifically warns about generating many pages without value, and calls for checking AI-generated content before publication. [8]

I would therefore avoid building a strategy around evading an imagined “AI-written” label. These reports reveal no such label. The stronger question is whether I can stand behind the facts, explain how they were checked, and give a reader something they could not get from another generic summary.

The advantage I want to build is a source that stays useful when an assistant needs an answer.

What comes nextThe recommendation gets more specific

I expect assistants to become better at working with a person’s actual constraints. A request for a barber can include a neighbourhood, a budget, a language, a preferred style and a time slot. A request about moving to Norway can include someone’s situation and the step they are trying to take next.

In that setting, the useful business information becomes more detailed. A clear price, an accurate description of a service, a working booking flow or a carefully sourced explanation may help an assistant decide whether a result fits. That is my expectation about the direction of discovery, rather than a ranking formula established by this case study.

I also expect comparison and verification to matter more as the volume of available content grows. But better models do not guarantee correct recommendations. Assistants can miss a source, repeat outdated information or sound more certain than their evidence allows. Independent corroboration and clear provenance remain valuable precisely because the process is imperfect.

This gives me a practical way to think about GEO, or generative engine optimization: the work of making a project discoverable and useful within AI-generated answers. It extends the questions I ask of SEO. Can the page be found? Can its claims be understood and checked? Does it help with this person’s particular problem? Does the product deliver when the person arrives?

SamfunnPrep’s results give me a reason to keep investing in that work. They do not establish a universal playbook, and updating a page or requesting indexing cannot guarantee continued growth. What they show is that a focused multilingual resource can already attract search visits and appear repeatedly as a source in AI answers.

The next version of search is still taking shape. My bet is that the work worth doing now is the work I would want an assistant to find later: specific answers, reliable sources, current information, and a product that deserves the recommendation.