Ask ChatGPT who builds AI agents for small companies in Europe and you get five names. Yours is not among them, though you have a site, a blog and valid structured data. That is not a bug, and another post on your own domain will not fix it.
An analysis of 23 387 citations on brand queries shows what those answers are built from: 57% rested on reviews and social proof, 17% on directories, and companies' own commercial pages accounted for 12%. Your own domain is one eighth of that material.
One caveat first. This is a field where certainty sells and evidence is thin. A survey of 45 GEO studies from 2023 to 2026 published on arXiv found no technique with a stable, cross-platform causal effect. Only one thing is graded moderately to strongly supported: extractable evidence, meaning statistics, comparisons, prices and dates. The same survey takes apart the popular "GEO lifts visibility by 40%" line: a relative maximum on one metric in a simulator with five documents.
Citation and recommendation are two different things
In a July 2026 study of 278 B2B SaaS prompts (3 508 citations, 2 845 unique pages, US and UK), 66% of brand recommendations happened without a single citation of that brand's website. The model named the company and linked to nothing of theirs.
Two mechanisms worth separating for good:
- Citation is the links under the answer. A specific page wins it because it matches the query and can be extracted.
- Recommendation is the sentence "you could ask company X", assembled from what the model knows about the market: third-party rankings, reviews, mentions, directories.
Your own site closes the first mechanism. The second plays out on other people's domains, so until you are there you can publish weekly and still miss the list.
What actually carries weight
Honest framing: these are correlations, not causation. The ranking is still instructive.
- presence in third-party "best of" lists: r = +0.64
- brand mentions on the web, including unlinked ones: r = +0.58
- Reddit presence: r = +0.57
- review count on G2: r = +0.39
That last number is awkward for anyone spending the whole budget on reviews in one platform. A separate dataset points the same way: brand mentions correlate with AI Overview citations at 0.664, classic backlinks at 0.218. In an analysis of 153 425 citations from May 2026, 76.95% of cited URLs sat outside Google's top ten, the exact area conventional SEO fights over.
Format matters too. Wix Studio AI Search Lab analysed 75 000 AI answers and over a million citations (March 2026): ranked listicles took 21.9% of all citations and 40% of commercial-intent citations, articles 16.7%, product pages 13.7%. Between 71% and 86% of cited listicles were numbered "10 best" formats with a comparison table and stated criteria.
And the fact that reorders how you think about review platforms: in the B2B study, G2, Clutch and Capterra pages themselves collected 1.6% of citations. They are not cited. They are read earlier and shape what the model says.
Directories and platforms worth the time
The "how it works" column matters more than the cost column: half of these cost nothing but hours.
| Platform | Who it is for | How it works | Entry cost | Time to fill in |
|---|---|---|---|---|
| Google Business Profile | any company with an address or service area | recommendation in local answers; GBP signals are about 32% of controllable local ranking weight | 0 EUR | 2-3 h plus photos |
| Clutch | B2B services: IT, marketing, implementation | verified reviews plus a marketplace app inside ChatGPT since 12 May 2026 | profile 0 EUR, paid packages | 3-4 h plus collecting reviews |
| DesignRush | agencies and studios | editorial "Best X" lists that themselves get cited in Perplexity and AI Overviews | profile 0 EUR, paid featuring | 2 h |
| GoodFirms | software houses and contractors | second-source corroboration when a model compares vendors | 0 EUR | 2 h |
| G2, Capterra | products and SaaS | weakest of the studied correlations with being recommended (r = +0.39) | 0 EUR | 2 h |
| Wikidata | any company with a web footprint | entity with no notability threshold; entity recognition shifts in 30-60 days | 0 EUR | 1-2 h |
| experts.n8n.io | n8n automation builders | partner directory; programme closed with a waiting list since mid-2025 | 0 EUR | 30 min to apply |
| technical and product topics | r = +0.57 for recommendation; 2.4% of ChatGPT citations, about 31% of Perplexity citations are social | 0 EUR, spam costs reputation | months of presence | |
| LinkedIn company page | everyone | 0.39% of ChatGPT citations, but a steady source of consistent company data | 0 EUR | 1 h |
General national business directories go last: they tidy up registration data, but no study I have seen gives them weight in model answers. Housekeeping, not a channel. For a company with a service area, Google Business Profile plus reviews carry most of it; I went through that in the piece on how a construction company gets clients online.
Clutch inside ChatGPT: what changed in May 2026
On 12 May 2026 Clutch launched a marketplace app that runs inside ChatGPT. The directory stopped being a badge and became a distribution channel: someone asking for a contractor gets a shortlist built from Clutch data without leaving the chat.
Keep the 1.6% figure in mind: the value is not in someone clicking your profile, it is in your data sitting in the set the shortlist is built from.
What the profile needs: services named in the client's words, a minimum project budget, an hourly rate, your industries, and reviews with specifics instead of praise. If you work alone, say one person. These forms are built for agencies and tempt you into inventing a team, the shortest route to a client who expects a department.
A profile and a site that can be quoted
Extractable evidence is the only thing with solid support in the research: numbers, prices, dates, comparisons, a source. Nobody quotes "projects at attractive prices". "AI visibility audit from 900 EUR (3 900 zł), two weeks" can be quoted.
The second point is counterintuitive for anyone selling: pages naming six or more brands averaged 2.13 citations against 1.21 for pages naming none. An honest comparison with Tidio, Intercom Fin, Make, Zapier, n8n or Voiceflow works in your favour: the model pulls you into the same context as recognisable names. Silence about competitors is a cost.
Third: freshness. The median age of a ChatGPT-cited page is 3.9 months, and 69.7% were published within the last twelve months. The update date has to be visible, and a 2024 price list does more harm than good.
Fourth is technical and most often neglected. AI crawlers do not execute JavaScript: as of June 2026 neither GPTBot nor ClaudeBot nor PerplexityBot does. If your price table or FAQ loads after hydration, it does not exist for them. I check that first when I take over someone else's company website.
One identity in every place
A model merges mentions into one entity by name, description and contact details. One legal name in the register, a second on the site, a third on Clutch: instead of one recognisable entity you have three weak ones.
I do it the same way every time: the same full name and one-sentence description everywhere, the same address and phone in the same format, sameAs in the structured data pointing at every profile, an entity in Wikidata. Wikidata has no notability threshold, and shifts in entity recognition are reported within 30-60 days of the item being created.
Structured data itself is plumbing, not a lever: an Ahrefs experiment on 1 885 pages found that adding JSON-LD moved citations by -4.6% in AI Overviews, +2.4% in AI Mode and +2.2% in ChatGPT, statistically nothing. More detail in the piece on what Schema.org still does and what quietly stopped.
Getting into someone else's list without paying for the slot
Third-party rankings show the strongest correlation with being recommended, so this is work worth doing by hand.
- Ask ChatGPT, Perplexity and Google AI Mode about your service and your city. Write down every list the model draws on.
- Check whether the author accepts submissions. Plenty of these roundups are one person refreshing a page quarterly.
- Send facts, not a pitch: what you do, for whom, starting prices, three projects with numbers, contact details. The author needs a table row.
- Publish your own roundup and name your competitors in it. It gets cited on its own, and gives someone a reason to name you back.
Paid placements exist and mechanically work, but that is rent, not an asset. Stop paying and you disappear.
What not to do
Three shortcuts that cost more than they return.
Manufactured Reddit activity. In Q1 2026 Reddit's automated systems removed roughly 25 000 spam posts a day, cutting user exposure by about 20% year over year. For a company trading under a named founder, the risk is out of proportion to the gain.
Betting on llms.txt. An Ahrefs study of 137 000 sites found 97% of llms.txt files received zero traffic in May 2026, and no major model provider has committed to supporting it. The file takes five minutes, but it is not a strategy.
Buying reviews. Verified platforms catch it more reliably every year, and one retracted review costs more than ten added ones build.
A one-hour monthly review
One hour a month is enough. I run the same five prompts in ChatGPT, Perplexity and Google AI Mode: "who does X in [city]", "best X companies in [country]", "X or Y, which to pick", "how much does X cost", "alternatives to [competitor]". I log who was named, which sources the model showed and whether a new list worth joining appeared.
Two anchors keep expectations sane. Google AI Mode has been live in Poland since 8 October 2025, Senuto puts an AI answer on 24.17% of Polish queries, and where one appears click-through drops by about 34.5% (Ahrefs, 300 000 keywords). Across roughly 350 000 locations studied, ChatGPT recommended 1.2%, Gemini 11% and Perplexity 7.4%, against 35.9% in Google's local pack. The recommended list is short, so the realistic goal is one narrow category, not "being everywhere".
If you want this on paper, I run an AI visibility audit from 900 EUR (3 900 zł): the same prompts across three systems, an inventory of the sources they pull from, directories and roundups ordered by impact, and a check of what AI crawlers actually see on your site. Prefer to do it yourself? The table above covers the first quarter, and if questions come up, write to me. The running order right after launch is in the piece on the first 90 days after a website goes live.
FAQ
Why does ChatGPT recommend a company but not link to its website? Because citation and recommendation are separate mechanisms. In a July 2026 study of 278 B2B prompts (3 508 citations, 2 845 pages), 66% of brand recommendations happened with no citation of that brand's site. The name comes from third-party rankings, mentions and directories; the links under the answer come from live retrieval.
Which directories are worth it for a B2B service company in Europe? Google Business Profile, Clutch and DesignRush first, then GoodFirms and a Wikidata entity. G2 and Capterra only make sense if you sell a product: G2 review volume correlated with being recommended at r = +0.39, the weakest studied signal. General national directories tidy your company data, but no study assigns them weight in model answers.
What did the Clutch app inside ChatGPT change? Clutch launched a marketplace app inside ChatGPT on 12 May 2026, so someone asking for a contractor can get a shortlist built from Clutch data without leaving the chat. A complete profile with reviews, budget ranges and industries becomes an acquisition channel, not a badge on your site. Review platform pages themselves took only 1.6% of citations in the B2B study: the effect runs through recommendation, not links.
Is it worth collecting reviews if review platforms are barely cited? Yes, but for a different reason than most people assume. Reviews and social proof were behind 57% of answers in the analysis of 23 387 citations on brand queries, directories behind 17%, own commercial pages behind 12%. Models read them when judging a company even if no link is shown. A sensible target is a dozen specific reviews on two platforms rather than a hundred on one.
How much does it cost to sort this out? Filling in the directories costs only time: 13 to 17 hours for the full set of profiles in the table above. Money starts with featured packages on Clutch and DesignRush and sponsored slots in third-party rankings. My AI visibility audit with an action plan starts at 900 EUR (3 900 zł) and includes a baseline, so there is something to compare against a quarter later.
Is Reddit worth the effort for a small B2B company? It depends who you sell to. Reddit presence correlated with being recommended at r = +0.57, yet it is only 2.4% of ChatGPT citations, while about 31% of Perplexity's citations come from social sites. For most B2B sellers it is a secondary channel, and manufactured activity is risky: in Q1 2026 Reddit removed roughly 25 000 spam posts a day.
