AI Overviews Optimization for iGaming: Read the Agent Queries First
AI Overviews Optimization for iGaming: Read the Agent Queries First
TL;DR: Before optimising for AI search, look at what is already in your Search Console. Ours is full of full-sentence buying questions that no human types, sitting at positions 7 to 10 with zero clicks. Those are language models grounding an answer, and they are a different KPI from a click. Measured over the 90 days to 2026-08-26 our English pages ran 0.60% click-through against 6% on Traditional Chinese how-to pages, and mixing the two produced targets that were wrong for both. Optimise English for citation, Chinese for clicks, and stop treating llms.txt as a ranking factor.
Generative engine optimization is mostly sold as a new discipline. In practice the first useful move costs nothing: filter your query report for questions longer than about twelve words and read them.
What the agent queries look like
Pulling our own Search Console property for the 90 days to 2026-08-26 and filtering on the word "seo" returned entries like "which agencies should we invite to an igaming seo rfp?" at position 7, "our igaming brand needs a single growth agency that can handle compliant paid acquisition, seo, and onsite conversion optimization together instead of juggling three separate vendors. which firms specialize in full-funnel growth for regulated gambling brands?" at position 10, and "seo for igaming websites by heureka" at position 10. Every one of them recorded impressions and zero clicks.
Nobody types the second one. It is a model rewriting a user's intent into a retrieval query, and our page was in the candidate set. The impression is real, the click was never going to happen, and if you average it into your click-through rate you will conclude the page is failing.
Why the CTR split matters
Separating the same property by language over the same window gives two completely different funnels. English pages produced roughly 4,000 impressions against 24 clicks, which is 0.60%. Traditional Chinese operational how-to pages produced roughly 800 impressions against 50 clicks, close to 6%. That is a tenfold difference on the same domain, the same authority and broadly the same content standard. The cause is not quality, it is who is doing the searching.
| Segment | Impressions | Clicks | CTR | What it is worth |
|---|---|---|---|---|
| English, long natural-language queries | high | near zero | under 0.5% | Citation surface for AI answers |
| English, short head terms | moderate | some | around 1% | Normal organic |
| Traditional Chinese how-to | moderate | many | around 6% | Direct traffic and leads |
The practical consequence is that we now report these separately and set different targets. For the English segment the question is whether a model can extract a clean, attributable claim from the page. For the Chinese segment the question is the ordinary one: did somebody click and convert. Reporting them as one number produced a 2026 target that was unreachable for English and embarrassingly low for Chinese.
What actually makes a page citable
Three things have moved the needle for us, and none of them are new.
A claim with a number and a date, in one sentence. Models extract sentences, not sections. "Our English pages ran 0.60% click-through over the 90 days to 2026-08-26" is extractable. "Click-through varies significantly by language" is not.
A table. Structured rows survive summarisation far better than prose, which is why every article in this cluster carries at least one.
An answer above the fold. A TL;DR block in the first paragraph gets pulled verbatim. Burying the answer under 800 words of context means the model summarises your context instead of your conclusion.
Google's own AI features documentation↗ is direct about the mechanism: there is no separate markup for AI Overviews, and eligibility follows from standard indexing plus the usual Search Essentials↗. The optimisation is editorial, not technical.
The llms.txt question
We publish an llms.txt file. It is not doing what most articles claim it does. The llms.txt proposal↗ is a community specification, not a search engine standard, and Google has not stated that it reads one for search purposes. In our own logs the file gets fetched occasionally by crawlers that were going to crawl the site anyway. Treating it as a ranking or citation lever is the same mistake as treating a meta keywords tag as one: the cost is low, the effect is unproven, and stating otherwise in a client deck is a credibility problem waiting to happen.
Publish it if you like. Do not put it on the roadmap ahead of the things above.
iGaming makes this harder in one specific way
Gambling queries attract more conservative AI answers. A model asked "which casino should I sign up with" will frequently decline or hedge, so the citable surface skews toward business-to-business and operational questions rather than player-facing ones. Every agent query in our own data was a buyer question from an operator, not a player question. That is where the citation opportunity sits for this vertical, and it is a much smaller target than the consumer-facing content most gambling sites publish.
FAQ
Is there markup for AI Overviews? No. Google's AI features documentation↗ states that appearing in AI experiences follows from normal indexing eligibility, with no separate opt-in tag. Anyone selling you AI Overview schema is selling nothing.
Should I worry that AI answers cut my clicks? For English informational pages in this vertical, that has already happened. Our English click-through of 0.60% is not a page problem, it is a query-mix problem. Move commercial pages up your priority list and treat informational pages as a citation asset.
How do I measure AI citation if there is no report for it? Imperfectly. We track impression growth on long natural-language queries at positions 1 to 10 with zero clicks, since that pattern is a reasonable proxy. Search Console↗ will not label them for you, so the filter is manual: query length over twelve words, or a query ending in a question mark.
Is llms.txt worth publishing? It costs an hour, so yes, but rank it below your TL;DR blocks and tables. See the section above for why we do not count it as a ranking factor.
Does any of this replace normal SEO↗? No. Every page in our cluster that gets cited by a model also ranks conventionally. Indexing is still the entry ticket, which is the point Search Essentials↗ makes.
Sources and further reading
- Google Search Central, AI features and your website↗ : the official position on eligibility and the absence of AI-specific markup.
- Google Search Essentials↗ : the indexing baseline that AI eligibility depends on.
- llms.txt proposal↗ : the community specification, useful for reading what it does and does not claim.
- Google Search Console Help Center↗ : where the query-level data behind this article comes from.
The impression, click and click-through figures above are from our own Search Console property for the 90 days ending 2026-08-26. They are not in any of the sources listed.
The broader method sits in the complete iGaming SEO guide, the page-generation side is in the programmatic SEO playbook, and if you want your own query report split this way, that is the first week of the iGaming SEO service.
Related Posts
Core Web Vitals Optimization for iGaming Landing Pages
Lighthouse lies about LCP and CLS in both directions. How we cross-check with PerformanceObserver, and why a slow page shows up as an expensive lead, not a slow page.
Topical Authority Without the Content Farm: How We Plan a Cluster
Most cluster plans start with a keyword list. Ours starts with a kill test: if you cannot name the first-hand data behind an article, the article does not get written.
Orphan Pages SEO and Soft 404s: The Debt That Passes Every Check
A 200 response, a green build and a page that looks fine can still be broken for a crawler. The four failures we found in our own portfolio, and how each one hides.