SEO and GEO (Generative Engine Optimization): How AI Search Changes Optimisation

Editorial review: 24 October 2026. AI answers can satisfy an information need before a visit to the source website. That creates a practical question for publishers and marketing teams: which content still attracts visits, and what commercial value follows? A citation, a visit and a subscription are different outcomes. Treating them as interchangeable leads to misleading success reports.

Search engine optimization (SEO) improves the discoverability and presentation of pages in search. Generative Engine Optimization (GEO) describes work aimed at visibility in generated answers. It does not establish a universal formula for being cited. A useful approach combines accessible pages, clear evidence and measurement of actual outcomes. The available research supports concern about referrals; it does not justify declaring all organic traffic or every publisher business model obsolete.

What zero-click evidence actually shows

A zero-click search ends without a click to an external result. That definition alone does not reveal whether an answer was useful, whether a brand was remembered, or whether a purchase happened later. It also does not identify AI as the cause of every missing click.

Pew analysed March 2025 browsing by 900 US adults: traditional-result clicks occurred on 8% of visits associated with AI summaries versus 15% without. Results were reconstructed in April, not captured alongside every visit. This observational comparison cannot isolate AI’s effect from differences between searches. [1]

A separate preregistered field experiment with 1,100 participants found increased publisher click-through when AI Overviews and AI Mode were removed. Published as an August 2026 preprint, it provides experimental evidence within its setting, rather than a forecast for every website or market. [2]

The practical conclusion is narrower than “AI kills traffic”: referral risk deserves measurement. A drop on a particular site still needs investigation into search demand, rankings, seasonality, technical changes and content mix. Higher conversion among remaining visitors is a hypothesis to test, not compensation that can be assumed.

Four outcomes that need separate measurements

Visibility

A link or mention appears. This does not establish that anyone visited the website.

Referral

A click brings a visitor. Analytics sessions and platform click counts can differ.

Action

A newsletter signup, enquiry or paid subscription starts. Its definition and attribution window need to be explicit.

Business value

Revenue, retention and delivery costs determine whether those actions support the publication.

This is a measurement framework, not a guaranteed sequence. A person may encounter a source repeatedly, visit directly later, or never visit at all. Missing attribution should remain unknown rather than being assigned to AI by assumption.

Publisher economics: fewer visits do not tell the whole story

The Reuters Institute’s 2026 industry survey covered 280 digital leaders in 51 countries and territories. It describes differing positions for subscription publishers with strong direct audiences and advertising-dependent businesses. These are industry perspectives from a strategic sample, not proof that one model will survive and another will fail. [3]

For an advertising-funded site, page views and advertising yield matter. For a subscription publication, acquisition is only part of the calculation: cancellations, renewals and the cost of serving readers also matter. A specialist publisher may earn from research or events. None of these models automatically replaces lost search revenue.

A fictional calculation

The following figures illustrate the arithmetic only; they are not market benchmarks or a prediction. Assume 1.5 page views per visit and €8 in advertising revenue per 1,000 page views.

• 10,000 visits produce 15,000 page views and €120 in advertising revenue.
• 8,000 visits produce 12,000 page views and €96 at the same yield.
• At a 0.4% paid-subscription start rate, those visit totals produce 40 and 32 starts respectively.
• A rise to 0.5% would bring the smaller audience back to 40 starts. That improvement must be observed; it cannot be inferred from the traffic decline.

The advertising calculation is page views ÷ 1,000 × yield. Subscription starts are not subscription revenue or profit. The example explains why a single traffic or citation number cannot answer the business question.

Options depend on the publication

Original reporting can provide a reason to return, but costs time and money. Curation can also add value through selection, verification and explanation. A newsletter offers a route to an ongoing relationship; its value depends on active readership and subsequent actions. A paywall needs content worth paying for and a workable acquisition path. Licensing requires an actual agreement with defined terms. Blocking access creates neither a buyer nor an obligation to pay.

Practical GEO work without citation promises

Google’s current guide keeps SEO relevant, requires indexed pages eligible for snippets, and rejects a special AI schema requirement. Eligibility still does not guarantee inclusion. These are Google-specific rules, not a specification for every answer engine. [4]

An editorial review can check whether a page answers its stated question, shows how its conclusions were reached and distinguishes facts from interpretation. Clear headings and accurate summaries help readers navigate. Named authors and relevant primary sources make responsibility and evidence easier to inspect. Neither a summary nor an author profile proves that an AI system will select the page.

Discovery is another separate step. Sitemaps help search engines find URLs without guaranteeing indexing; IndexNow notifies participating engines about changes without guaranteeing inclusion. Neither mechanism promises an AI citation. “Real-time indexing” is therefore an unsuitable blanket description. [10] [11]

Search visibility, training and access are separate decisions

Google’s Search Console control allows inclusion in or exclusion from specified Search generative AI features; inclusion is the default, subject to property inheritance. Google states that exclusion does not act as a ranking signal for other Search areas and does not govern AI training. [5]

Google-Extended covers the documented Gemini training and grounding uses; it is not the switch for inclusion in Google Search. Grounding means supplying source material when an answer is generated. Training and answering with retrieved material must not be conflated. [7]

A robots.txt rule relies on crawler compliance and is not access control. Sensitive or paid content needs appropriate server-side protection where access must actually be restricted. [8] A new crawler rule also provides no evidence that previously acquired material has been deleted or removed from an already trained model.

A measurement plan that can support decisions

Google’s documented generative AI performance report shows impressions for AI Overviews and AI Mode, with page, country, date and device breakdowns. Its documented metric is impressions, not a complete AI referral or revenue report. The dates use Pacific Time. [6]

Reading the report: An impression records a displayed site link, not necessarily a visit or an unlinked brand mention. Too few impressions can prevent the report from appearing. Page data generally uses canonical URLs; property totals and summed page rows can differ. These details need to match the CSV comparison rules. [6]

1. Define a question, such as whether informational pages lost referrals while their search exposure remained stable.
2. Record the source, period, time zone, URL rules and available dimensions of each dataset.
3. Compare similar page groups and periods; record editorial releases and tracking changes.
4. Evaluate sessions, meaningful actions and business results separately from visibility.
5. Document alternative explanations. A before-and-after change alone does not prove that GEO work caused a result.

The AI citation and website traffic comparator places prepared citation and session CSV data alongside each other by URL. It does not connect individual citations to visitors or calculate a valid click-through rate from those two totals.