Getting named by AI assistants in a category where the incumbents own the citations
Kestrel Financial Services was almost never named by ChatGPT, Gemini, Claude or Perplexity for the questions its buyers asked, and reached consistent citation over seven months by resolving its entity, restructuring content for extraction and earning corroboration from sources the models already trusted.
Measured outcomes
- Citation rate across the panel
- 5% → 47%
- 80 prompts, four assistants, month 1 vs month 7
- Assistants naming Kestrel at least once
- 1 of 4 → 4 of 4
- Sustained across the final three monthly runs
- Correct-attribution rate
- 38% → 89%
- Share of citations describing Kestrel's category accurately
- Referral sessions from assistant surfaces
- 0 → 1,340/mo
- Month 7; small in absolute terms and growing
The situation
Kestrel ranked respectably in classic search but was effectively invisible in generated answers. A baseline panel of eighty buyer questions returned its name four times. Three competitors appeared in more than half.
The entity itself was ambiguous. Kestrel shared a name fragment with an unrelated asset manager and a defunct insurance brand, and its own properties disagreed about what it did: the website said payments infrastructure, the Crunchbase entry said lending, and the LinkedIn description said neither.
The content was written to persuade a reader already on the page. Claims depended on surrounding context, comparisons were implicit, and almost nothing survived being lifted out of its paragraph, which is precisely the operation a retrieval system performs.
What we did
Baseline before touching anything
We built an eighty-prompt panel from the questions the sales team actually fields, and ran it across four assistants on a fixed schedule. Without that baseline every later claim about improvement would have been unfalsifiable, including ours.
Resolve the entity
One canonical organisation node with a stable identifier, consistent naming and description across every controlled property, and explicit disambiguation from the similarly named firms. This is unglamorous work and it was the single highest-impact change in the engagement.
Rewrite for extraction, not for flow
Key claims were rewritten to be self-contained: each opens by naming its subject, states the claim in one sentence, and carries its own qualifier. We rebuilt the comparison content to state positions explicitly rather than implying them, because a model summarising an implication usually gets it wrong.
Earn corroboration off-site
Models weight sources they already trust, and none of that sits on your own domain. We ran a targeted programme into industry publications, standards bodies and two academic collaborations. This is the slowest component and the one clients most often want to cut. We recommended against cutting it.
What we would do differently
Correct attribution mattered more than citation frequency and we did not weight it heavily enough at the start. Early in the engagement Kestrel was being named more often but described as a lending platform, which generated unqualified enquiries and briefly made the sales team's experience worse, not better. We now track attribution accuracy from day one.
“For the first two months the honest answer to 'is this working' was 'we do not know yet, here is the panel'. I would rather have that than a dashboard that always goes up.”