AI search optimisation for R&D tax credit specialists: what actually gets a firm recommended
One firm was cited 36 times in our sweep of UK R&D tax advisers and recommended not once. Another converted 62% of its mentions into the specific pick. Same questions, same week — the difference was what the engines could point to as the reason.
Steen Stones · Reviewed 5 Aug 2026
If a business asks an AI who should handle its R&D tax claim, the engine has an opinion — and in our own sweep, that opinion did not track with who gets cited most. One firm was named 36 times across the sweep, more than most firms in the tally, and recommended in precisely none of those answers. EmpowerRD, named 27 times — fewer — converted 15 of those into a specific recommendation. Same market, same 50 questions, same week.
Avouch multi-engine market sweep, 18 July 2026 — 50 buyer questions, UK R&D tax credit advisers, ChatGPT + Gemini + Claude, 87 firms tallied
That gap matters more here than in most sectors, because HMRC has spent the last two years tightening scrutiny on who it will accept an R&D claim from. A firm an AI cites constantly but never puts forward as the answer is, commercially, close to invisible — clients don't act on a name buried in a list, they act on the one the engine tells them to use.
Named is not chosen
Below is what the conversion from named to recommended actually looked like for a representative set of firms in our sweep — comparable mention counts, very different outcomes. The firms shown converted well; the ones that converted poorly are described, not identified, because the pattern is the point and their identity is not.
number of AI answers each firm appeared in · named = mentioned anywhere · recommended = put forward as the pick
ForrestBrown converted 62% of its mentions into recommendations. EmpowerRD and Kene Partners, named at roughly the same rate as each other, converted 56% and 33% respectively. Nothing about the questions changed between them.
What the engines actually cite as the reason
Read the recommending sentences and the pattern is specific, not generic — the same shape we found in the solicitors sweep. Claude, asked which firm handles claims for a biotech startup:
“My genuine recommendation: Kene Partners for an early-stage biotech startup, because PhD-level technical understanding combined with tax expertise reduces the risk of HMRC challenging the science behind your claim.”
Two more from the same sweep: asked which firm has the strongest compliance record since HMRC's crackdown, Claude named RCK Partners specifically for its “extra lawyer-reviewed sign-off layer”. Asked about contingency-fee providers, it picked RandDTax as “an established, fee-competitive contingency provider with a long track record — average claim value over £145,000 — rather than one of the newer volume-driven players”. In every case: a named capability, a stated reason, a number where one exists. Not “a well-regarded firm”.
Compliance credibility is the wedge, not a footnote
HMRC's own compliance activity shapes this market more directly than in most sectors. The Additional Information Form became mandatory for every R&D claim from 8 August 2023, and the SME and RDEC schemes merged into a single scheme for accounting periods starting on or after 1 April 2024. Our sweep shows that credibility signal already being used as a selling point in the market — one firm's own sponsored search copy leads with “HMRC-Ready R&D Tax Claims, under 5% HMRC enquiry rate” rather than price. The firms converting best in our data lean on exactly this kind of specific, checkable compliance claim, not a generic “we're experts” statement.
Every engine behaves differently — and one behaves the same way it did in a completely different market
Across all 87 firms in this sweep, Gemini's recommendation count is zero. Not low — zero, for every single firm, across all 50 questions. We found the identical pattern in a separate sweep of Southampton solicitors two weeks earlier: Gemini names constantly, and in both markets we've checked so far, never once puts a specific firm forward as its pick. ChatGPT and Claude, asked the same questions, both converted real — and different — shares of their mentions.
That is the clearest evidence yet for a rule that applies well beyond R&D tax: measure recommendation rate per engine, never as a blended average. A firm doing everything right could show a flat, discouraging number if Gemini's zero is folded into the same total as ChatGPT's real conversions.
“There's a difference between showing up and being recommended. AI cannot recommend what it doesn't know.”
What this means for a UK R&D tax adviser
- Lead with a specific, checkable compliance claim — enquiry rate, sign-off process, sector specialism — not a general expertise statement. It's what every recommending sentence in our data actually cites.
- State a real number where you can back it up. Average claim value, years of track record, enquiry rate — vague superiority claims are unusable to an engine; specific figures aren't.
- Don't confuse citation count with commercial outcome. Being the most-named firm in a market meant nothing here without a specific reason attached to the name.
- Track recommendation rate per engine. Our data shows one major engine converting zero across two entirely separate markets — folding that into an average hides real progress on the other two.
Common questions
- Is R&D tax genuinely different from the law firm sweep, or is this the same finding repeated?
- Both. The mechanism is identical — named is not chosen, specific claims convert, engines behave differently from each other — which is exactly why it's worth publishing twice: two unrelated UK professional-services markets, swept two weeks apart, producing the same shape of result independently. What's different here is the specific evidence buyers respond to: HMRC compliance credibility and claim-value numbers, not directory rankings.
- Why is Gemini's recommendation rate zero across two whole markets?
- We can see the pattern clearly but not the cause from this data alone — Gemini appears to default to listing options for this kind of comparative professional-services question rather than picking one, at least in the sweeps we've run. It's exactly why every figure here is reported per engine rather than blended into one score.
- Why won't you name the firms that converted poorly?
- The pattern is what matters, not any one firm's identity, and quoting public material about a named firm's poor conversion carries legal risk with no offsetting benefit here — the finding holds just as well described as it would named.
AI can't recommend what it doesn't know.
Start free. No commitment.