Study
Which IP firms do AI assistants recommend?
Once a month, the same questions are put to four AI assistants for every country IPRATE publishes, and the firms they name are compared with the register-derived ratings. The panel is fixed, the answers are stored, and the per-query dataset is open.
Issues
The first issue is in preparation. It covers the month just ended and is published on the eighth of the following month.
Method
- A fixed panel. Three questions per country, in English and in the country's own language, plus twenty cross-cutting questions. Rewording a question is a version bump, and every issue prints the version it used, because a reworded panel makes two months incomparable.
- Official APIs only. No consumer interface is scraped. That respects the vendors' terms and it is the only way the run can be repeated by somebody else.
- Three repetitions. The same question is asked three times per assistant per month, so a one-off answer shows up as instability rather than as a finding.
- Stored, then read. The raw answer is the record. Names are extracted from the stored text afterwards, so the extraction can be corrected and rerun without querying anything again.
- Named only when published. A firm or attorney appears by name only where it holds a published IPRATE profile. Everything else is counted.
What this study is not
It does not feed ratings, profiles or search. No IPRATE profile shows whether an assistant recommended it, and nothing an assistant says changes a score, a tier, a rank or a confidence grade. It is a measurement of where assistants currently send people, held next to what the registers show, and nothing more than that.
It is also not an audit of any assistant. A name IPRATE cannot match may be a firm IPRATE does not hold, a firm under another name, or a mistake. The study reports which, where it can tell, and says so where it cannot.
Questions
Which assistants are measured?
OpenAI, Perplexity, Google Gemini and Anthropic Claude, each through its official API with web search enabled. Microsoft Copilot has no public API and is excluded rather than scraped; every issue says so, so its absence is not read as a zero.
Why is Gemini used as the proxy for Google AI Mode?
Google AI Mode has no API. Gemini with Search grounding is the closest queryable surface. It is a proxy and the issue states that plainly; it is not the same product and the numbers should not be read as if it were.
What happens to a name the study cannot match?
It is counted as unmatched and kept verbatim in the open dataset. An unmatched name is not thereby incorrect and this study does not call it fabricated: it means only that IPRATE holds no entity it can be matched to with confidence.
Does being recommended by an assistant affect an IPRATE rating?
No. The study reads published outputs and writes nothing back. No profile shows whether an assistant recommended it, and nothing an assistant says can change a score, a tier, a rank or a confidence grade.
Can the study be reproduced?
That is what it is for. The query panel is fixed and versioned, every issue prints the panel version, the model ids and the dates, the raw answers are stored, and the per-query dataset is published as CSV under CC BY-NC 4.0.
The ratings the study compares against are described in the methodology. The monthly filing figures are the market overview.