Talking Tom has billions of downloads. AI still did not recommend it — what a Slovenian studio did in six months

Why does this case matter to a company that does not make games?

Because it is the cleanest public example we have found of the problem we measure, written by the company itself rather than by an agency selling a fix. Outfit7 — the studio behind Talking Tom & Friends, founded in Ljubljana in 2009 — published a first-person account on 1 September 2026 of how it discovered that AI engines did not recommend its games, what it did about it, and what changed.

The mechanics are the same for an accounting firm in Ljubljana or an industrial supplier in Maribor. Only the scale differs — and that is what makes the case instructive: if billions of downloads do not earn a recommendation, a good reputation in your city will not either. The evidence has to be where the engine looks.

What did Outfit7 find when it asked the engines?

In October 2025 the company audited how the major AI engines described and recommended it. Two kinds of questions produced two different pictures.

Asked directly — "Who is Outfit7?" — the engines answered accurately, drawing on the company's Wikipedia page: founders, scale, corporate history. Asked the way a player asks — "I'm bored, what virtual pet mobile games would you recommend?" — the engines named smaller, niche competitors. Talking Tom, which the company says defined the modern virtual-pet genre, was not in the answer.

This is precisely the distance between step two and step one of our check: the engine knows you, and still does not choose you. Outfit7 had a positioning problem, not an awareness problem, and the two need different work.

How did they measure it without any data from the AI companies?

They could not. Search engines give site owners a console; AI engines give nothing. Outfit7's answer was to build the buyer's side of the conversation from the outside in:

From this they wrote a master list of 50 prompts and used it as a fixed baseline: the same prompts, re-run across engines, with presence in the answer recorded each time. They also tracked how many visitors arrived at their sites from AI assistants. Two proxy numbers, chosen because the real data is locked away — and both repeatable.

What did they change on their own sites?

Four things, none of them exotic:

1. Named the entity. "Our studio's premium products" became "Outfit7" and "Talking Tom & Friends" so that models could index who was being talked about.
2. Wrote in questions and answers. Landing pages got subheadings phrased like real prompts, each followed by a short, standalone answer.
3. Made the structure machine-readable. Bullet points, short paragraphs, comparison tables; JSON-LD structured data for the organisation and its products; registration in Bing Webmaster Tools.
4. Kept publishing. Dedicated pages per game — gameplay, milestones, awards — then regular guides and feature explainers to widen the topical footprint.

The direction matches the only controlled study of generative engine optimisation we know of: pages with citations, statistics and quotations were used more often in generated answers, and keyword stuffing was used less. Outfit7 made its pages look like evidence.

What did they change outside their sites?

The company's own conclusion was that engines "prioritize neutral collective wisdom over a brand's own marketing copy". So the second half of the work went to third parties:

• Press releases rewritten with a plain summary at the top, so that the facts could be extracted without reading the rest.
• Editorial pitching to the high-authority sites engines cite repeatedly — Outfit7 calls them "super feeders".
• Placement in "Top 5" and "Top 10" lists, because a recommendation question is answered from recommendation lists.

For a B2B company the equivalents are the industry directory, the trade publication, the chamber of commerce listing and the review platform — the places an engine reads when a buyer asks "which firms should we talk to".

What changed after six months?

By Outfit7's own account, two numbers moved:

• Presence in about 90 % of the 50 tracked prompts, from prompts where the brand had been "completely invisible".
• +357 % visitors arriving through AI chatbots and assistants since the work began in October 2025 — roughly four and a half times the starting level.

The company also says the larger gain was understanding how the answers are assembled, and that it now applies the same rules to daily PR and web copy.

What we did not verify

The results are self-reported. Outfit7 has not published the prompt list, the engines, the run dates or the starting traffic figure, and no independent party has repeated the measurement. We quote the numbers because the method is sound and the source is first-party; we do not treat them as a benchmark for anyone else. Our own checks store engine, model, date and sources for every answer so that a third party can repeat them — the standard we would ask of any case, including this one.

What should a B2B owner take from it?

Three things. First, ask the two questions separately — about you by name, and about your category without your name — and do not let the first answer reassure you about the second. Second, write your buyer's prompts from what buyers actually say to you, not from a keyword tool. Third, put the same facts, in the same words, on your site and in the independent places the engines already trust — then measure again on a known date.