In one sentence: a model writing a recommendation prefers claims it can see corroborated by a source it treats as independent, so the highest-leverage work is often making a true claim about you appear somewhere you do not own.
Why the model weights it this way
This is not a ranking rule imposed from outside; it falls out of how the answer gets constructed. A claim sourced only to the vendor's own marketing page is the weakest kind of evidence available, and a system assembling an answer from multiple retrieved passages will naturally favour agreement between independent sources over emphasis within one.
Your own domain establishes what you claim. Everything else establishes whether the claim holds.
Which surfaces actually carry weight
Read your own citation data first — the source mix in answers to your question set tells you where the engine looks for your category, and it differs by category more than people expect. In general, ordered by effort against effect:
| Surface | Effort | What it corroborates |
|---|---|---|
| Directory and marketplace listings with every field completed | Low | Category, pricing shape, integrations, who it's for |
| Structured knowledge bases | Low–medium | That the entity exists and is distinct |
| Comparison and review pages you do not own | Medium | Position relative to named alternatives |
| Community threads where people describe the problem | Medium–high | Real usage, in the buyer's own vocabulary |
| Customer-authored documentation and case write-ups | High | Concrete capability claims with specifics |
The rule that keeps this honest
Everything here has an obvious abuse pattern, and the abuse pattern is also the failure pattern. Fabricated reviews, astroturfed threads and paid placements that read as organic are detectable, they poison the entity you spent months building, and — narrowly — they do not survive a buyer clicking through. The version of this work that compounds is slow: complete listings honestly, answer questions in communities where you have something specific to say, and let customers write what they actually experienced.
What to do when you have nothing yet
Start with the surfaces that require only accuracy, not persuasion. A directory listing with every field filled in, a knowledge base entry with the founding facts, documentation that states capabilities precisely including their limits. These are dull and they are the fastest corroboration available, because they need no one's permission and no one's enthusiasm.
How to tell it is working
Not by counting mentions. Track the gap between mention rate and recommendation rate over a fixed question set. Corroboration is the thing that closes that gap — a brand mentioned in 40% of answers and recommended in 4% is exactly the profile of a brand with layer one and layer two and nothing beyond its own domain.