Research
Experiments, data and research on how brands are discovered by AI.
The POI369 AI Search Lab is the research arm of POI369, a creative and technology agency in Tallinn, Estonia. It runs repeatable experiments on how AI answer engines (ChatGPT, Gemini, Perplexity, Google AI Mode) discover, understand and cite brands, publishes the method and the prompt sets openly, and reports results only after they have been measured. The first study is a monthly visibility baseline for POI369 itself.
Most claims about AI search visibility are opinions. We would rather publish numbers, including the ones that are inconvenient for an agency selling GEO. So the Lab has a simple rule: every study lists its prompt set, its engines, its scoring sheet and its date, and nothing is reported that was not measured.
The work starts small. Our own brand is the first subject, because we can observe every change we make to the site and watch what the engines do with it. Wider studies of Estonian companies and of specific page changes follow once the baseline method has run for a few months.
Last reviewed September 14, 2026 · Reviewed by Ahmet Bulut
One fixed prompt set, four engines, the same scoring every month. The set has three groups: brand prompts (the company name and its variants), commercial prompts (category phrases a buyer would type) and conversational prompts (the way people actually ask an assistant).
Share of answers in which the brand is named at all, per engine and per prompt group.
Share of answers in which a poi369.com URL is cited as a source. Mention without citation is recorded separately.
Where the brand appears among the recommendations in the answer: first, second, third, later, or only in passing.
Which other companies appear in the same answers. Over time this shows who the engines consider our peers.
The URL the engine cited when it mentioned us. The single most useful field: it shows which page, or which third-party site, the engine learned from.
The full set runs in the first week of every month, by hand, in a fresh session per engine, from Estonia. Results go into a shared sheet with the date and the engine version where visible. Bing Webmaster Tools AI Performance, Google Search Console and GA4 referral data are recorded alongside so the qualitative picture and the traffic can be compared.
Baseline run: September 2026. First comparison published after the second monthly run.
Fifty prompts (brand, commercial, conversational) run monthly across ChatGPT, Gemini, Perplexity and Google AI Mode, before and after the site’s GEO changes. Mention rate, citation rate, position, competitor overlap and citation source, published with the raw sheet.
After Study 01 has two data points.
A controlled comparison: one commercial page kept as is, one rebuilt with a direct-answer passage, evidence section, connected structured data and honest Q&A. Same prompts, same engines, same months. We report what moved and what did not.
Planned for late 2026.
A category-level look at how often Estonian companies are named and cited by answer engines for their own category prompts, which sources the engines rely on, and how entity consistency relates to visibility. Companies named only with their own public information.
Ongoing log analysis, first write-up when there is enough data.
Server-log analysis of AI and search crawler behaviour on poi369.com: which user agents fetch which pages, how often, whether they request the Markdown version, and how quickly changes are re-fetched after IndexNow pings.
Every study lists prompts, engines, dates and the scoring sheet so anyone can repeat it.
If a study has not run, it is listed as planned. Nothing is presented as a finding until it has been measured.
If the engines ignore us, the report says so. The Lab exists to learn, not to advertise.
Comparative studies use what a company has published about itself. No private data, no scraping against terms of service.
Between a company’s website and a sentence in a ChatGPT answer there are four stages: crawling, retrieval, entity resolution and citation selection. Each one can silently drop a brand. This is a plain-language map of the pipeline and of what a brand controls at each stage.
The concrete decisions behind poi369.com and the products we build: what goes in server-rendered HTML, how a page serves Markdown to agents on request, why sitemap dates are derived from content signatures, how six languages stay self-canonical, and how structured data forms one graph. Everything here is running in production.
We run the same baseline for client categories as the first step of every AI search engagement. Ask for a benchmark and we send the sheet and a plan.
