Getting Recommended by AI Assistants: Turning a Lucky Referral Into a Repeatable Channel

Whether an AI assistant recommends your brand is decided largely outside your own website: 85% of brand mentions in AI answers come from third-party sources, according to AirOps, and only 13.2% from the brand's own domain. The channel becomes plannable not through a single measure but through a control loop: measure a fixed prompt set, close the gaps in third-party sources, measure again. What you steer is the probability of a mention — not any individual answer.

The trigger for this article was a real search query that reached us: a company won its best new customer of the month through a ChatGPT recommendation. It was luck — and the question is now how to turn that into a plannable channel. That is what this article answers, including the criteria for evaluating a suitable provider.

Citation, mention, recommendation — what is the difference?

These three terms get mixed up regularly, even though they mean different things and require different work:

The case here is the third row: someone asked an assistant about providers, your name was in the answer, and that became a customer. Anyone who wants to repeat that effect has to understand that a recommendation is a different goal than a citation — and that you can be recommended without ever being cited.

Was that one recommendation luck?

Most likely yes, at least in part. AI answers are non-deterministic: Profound analysed more than 240 million ChatGPT citations and found that 40–60% of cited domains change from month to month for identical queries. SE Ranking ran 10,000 queries through Google's AI Mode three times: only 9.2% of cited URLs were identical across all three runs.

Before you build a strategy on it, run a simple test — half an hour of work.

1. Reconstruct the prompt. Ask the new customer how they searched, or derive the wording from Search Console. Use their phrasing, not a cleaned-up version of it.
2. Query it ten times, each in a fresh chat. No history, no personalisation, and no brand name in the prompt.
3. Test across several systems. ChatGPT, Perplexity, Gemini and Google AI Mode cite different sources. According to AirOps, 68% of brands appeared in only one AI platform — a result in one system says little about the others.
4. Count. Eight mentions out of ten means you hold a real position worth expanding. One or two means it was noise, and you are starting from zero.

Both outcomes are useful. The only difference is the time horizon: an existing position can be expanded in weeks, a new one takes months.

Which signals decide who an AI assistant recommends?

The most important and most uncomfortable finding first: most of the effect does not sit on your website. The AirOps analysis cited above — 21,311 brand mentions across 500 commercial-intent queries in six verticals, tested on GPT-5, Claude Sonnet 4.5 and Perplexity Sonar — shows that brands are mentioned 6.5 times more often through third-party sources than through their own domain, and that nearly 90% of those third-party sources are listicles, comparisons and reviews.

What does not show up is equally instructive. A study by OppAlerts covering 145 industries and more than 105,000 ChatGPT prompts, analysed by The Digital Bloom, finds only a weak correlation between backlinks and AI recommendations (Spearman rho +0.204, R² 4.2%). All 13 signals examined together explain well under 20% of the variance. Translated: classic SEO metrics do not buy you a recommendation, and anyone guaranteeing you a placement is selling something the evidence does not support.

What can be influenced comes down to six levers, ordered by effect on recommendation probability:

The last two levers are hygiene: they win you nothing, but their absence costs you everything. The firewall trap in particular is common — bot protection at the CDN layer often blocks AI crawlers by default without anyone noticing. How the content side works is covered in From SEO to GEO.

How does this become a plannable channel?

Through a four-step control loop run every quarter. The channel is created by the repetition, not by any single measure.

1. Build the prompt set (week 1). 50–150 realistic questions along genuine decision situations, structured by funnel stage: unbranded ("Who helps SMEs in Zurich with X?"), category ("Compare the best providers for X"), branded ("Is [your brand] reputable?"). The set stays stable across quarters, otherwise no comparable trends emerge.
2. Measure the baseline (week 2). Query each prompt repeatedly across all relevant systems and record per answer: mentioned yes/no, at which position, in what tone, and which sources the AI drew on. That last point is the most valuable — the cited domains are your action list.
3. Close the gaps (weeks 3–10). Work where the AI sources its answers: get into the relevant comparison articles and directories, build review profiles, unify entity data, and create content for the prompts where you are absent.
4. Measure again and report (weeks 11–12). Same set, same method. Report the change in mention rate and share of voice against the baseline. How such reporting is structured in detail is covered in Measuring AI visibility.

A realistic expectation: movement in the mention rate is measurable within one quarter, a stable channel takes two to three. That is no worse than other channels — it is the same horizon as SEO, with the difference that competition here is still thin. According to McKinsey, only 16% of companies measure their AI search performance systematically.

The traffic quality justifies the effort: visitors from AI search convert at 4.4 times the rate of classic organic visitors, according to Semrush. And the market is there: in Switzerland, 60% of the population already uses AI tools according to the IGEM Digimonitor 2025, rising to 79% among 15- to 34-year-olds.

How do you evaluate a suitable provider?

We deliberately do not name a list of providers here. The market is young, the labels — GEO, LLMO, AI SEO, AI visibility — are used inconsistently, and any list would be outdated in six months. Criteria you can test yourself in a first conversation are more useful:

1. An in-house measurement setup, not a resold tool. Ask how the prompt set is built and how the volatility of AI answers is handled. Anyone merely reselling a SaaS dashboard delivers data, not actions.
2. Work on third-party sources, not just your website. If the offer consists solely of on-page optimisation, it addresses 13% of the problem. Ask specifically how the provider builds presence in comparisons, directories and trade media.
3. Probability promises, not placement promises. "We'll get you to position 1 in ChatGPT" is not a credible promise for non-deterministic systems. Sound providers talk in mention rates and share of voice.
4. Measurement across several systems. Looking only at ChatGPT misses the larger part of the market — 68% of brands appear in a single platform only.
5. A connection to traffic and lead data. Visibility without a link to GA4 and CRM stays a vanity metric. Good reporting shows whether mentions turn into enquiries.
6. Demonstrable AI visibility of their own. Ask an assistant the provider question for your category. Anyone raising other companies' recommendation probability should be findable in their own.

The same logic in broader form — for choosing an agency beyond AI visibility — is covered in Choosing an SEO agency in Switzerland: 7 criteria.

What you can do yourself in the next 30 days

Even without external support, the starting position can be improved considerably:

• Run the luck test described above and write down the result — that is your baseline.
• Check crawl access: does your robots.txt allow GPTBot, ClaudeBot, PerplexityBot and Google-Extended? Does your CDN or firewall block AI bots?
• Unify entity data: write company name, address, services and founders identically on your website, LinkedIn, Google Business Profile and in directories; add Organization schema with sameAs.
• Read out the AI's sources: note which websites are cited in every test run. That list is your most important working document — it is where recommendations are decided.
• Collect five reviews on the platforms cited in your category.

These five steps cost little and remove the most common disqualifiers. Everything beyond that — prompt set, continuous measurement, systematic work in third-party sources — is project work.

Conclusion

The answer to the original question: yes, a chance AI recommendation can be turned into a plannable channel — as a probability, not as a placement. The path runs through a stable prompt set, repeated measurement across several systems and consistent work on the third-party sources models build their recommendations from. Optimising your own website alone addresses 13% of the problem.

Hierarchy runs its own in-house AI visibility tracking for exactly this: continuous measurement of brands across ChatGPT, Google AI Overviews, AI Mode, Perplexity and Gemini — with custom-built prompt sets, repeated sampling against answer noise and competitive benchmarks. Measurement and execution come from one team, as part of our SEO & content offering.

Won a customer through an AI recommendation and want to know whether it can be repeated? Let's talk.

Sources

• AirOps – The Influence of Offsite Signals in AI Search (October 2025) – 21,311 brand mentions across 500 commercial-intent queries: 85% from third-party sources, 13.2% from the brand's own domain, nearly 90% of third-party sources are listicles, comparisons and reviews; 68% of brands visible on one platform only.
• The Digital Bloom – LLM Ranking Factors 2026 – analysis of the OppAlerts study across 145 industries and 105,000 ChatGPT prompts: backlinks rho +0.204 (R² 4.2%), all 13 signals together explain under 20% of the variance.
• Profound – AI Search Volatility – 40–60% of cited domains change monthly.
• SE Ranking – AI Mode Research – only 9.2% identical URLs across three runs.
• Semrush – ChatGPT Search Insights – AI search visitors convert at 4.4x.
• McKinsey via MarketingTech – only 16% of companies measure AI search performance systematically.
• IGEM Digimonitor 2025 – 60% of the Swiss population use AI tools.
• Aggarwal et al. – GEO: Generative Engine Optimization (KDD 2024) – up to 40% visibility gain through quotable content optimisation.