A founder told us last month that his best lead of the quarter had opened with the sentence: “ChatGPT recommended you.” He had done nothing for it. The question is how to make that reproducible.
The shortlist is made in the answer
B2B buyers now ask ChatGPT, Perplexity or Gemini which vendors exist for a problem before they open a single website. The answer names two to five, with reasons. Whoever is on it gets a visitor who has already been advised. Whoever is missing is not compared, but never considered in the first place.
For startups this is an unusually fair starting position. The models do not favour the biggest brand but the best-evidenced one. A young company with clean sources can overtake the established vendors in a niche, without their budget.
A model recommends what it can verify in third-party sources. Your website is the claim, not the proof.
Sources instead of pages
The most common mistake: applying the SEO playbook to GEO. More blog posts, more keywords, more landing pages. But models do not rank pages. They cite sources they trust, and those are almost never on your own domain.
- Directories and comparison portals of the category (G2, Capterra, OMR Reviews, industry-specific lists)
- Trade press and guest articles that tie your name to the problem
- Reviews with text, not just stars
- Wikidata and company profiles: the entity “brand = category for audience” as a fact that can be looked up
- Communities where your buyers actually ask (Reddit, specialist forums, LinkedIn)
Your own website has one job: to answer the buying questions so directly that a model can cite the paragraph. FAQ schema, clear comparison pages, no marketing prose.
Which source types AI answers cite in B2B software
Entity instead of keyword
Before a model recommends you, it has to know you unambiguously. Two questions decide: does the model know which category you compete in? And does it confuse you with nobody? At one of our customers the brand was mixed up 54 times a month with a similarly named company from another industry. After three weeks of entity work it was three.
Concretely: a Wikidata entry with category, founding year and website; identical descriptions on LinkedIn, Crunchbase and in directories; Organization schema on the home page with sameAs references. That is one day of work and the lever with the best ratio of effort to effect.
The 90-day plan
Ninety days is not a marketing promise but the time it takes to measure, act and measure again. A realistic sequence for a team of two:
Baseline
+Define 20 buying questions, measure weekly
+Wikidata, profiles, Organization schema
+List in three directories of the category
Build
+Collect ten reviews with text
+One guest article in the trade press
+FAQ page with the 20 questions, FAQ schema
Proof
+Before/after per question
+Compare cited domains
+Decide what continues
The most important point is the measurement. Ask the same 20 questions every week on the same platforms and note who gets named. Without that series you will not know after 90 days whether anything worked.
Three mistakes we see again and again
- Content volume instead of source work. Twenty blog posts move less than two trade-portal articles and a clean review profile.
- Investor perspective instead of buyer question. The website explains the product; nobody answers the question the buyer actually asks.
- Measuring once. A snapshot is not a trend. Models do not answer deterministically; only the weekly average shows the direction.
What you can do this week
Ask ChatGPT and Perplexity the five most important buying questions of your category. Note who gets named and which domains get cited. If your name is missing, you now know why, and where to start.