“Wait, Is That URL Ours?”
Your company name comes up in an AI answer, and there’s a URL sitting next to it that you don’t recognize.
I had this happen last month: a subdomain I’d never seen turned up in an answer about us, and I couldn’t think of a single person to ask about it.
Campaign domains, domains from the old company name, one domain per product line. They scatter further than anyone expects, and in plenty of companies the only complete picture lives with IT.
I let it go with a “probably ours” — which wasn’t a judgement call so much as an admission that I had no way to check.
But there is an evaluation of how well models can tie a brand to its domains, published on 18 June 2026, and it turns that “probably” into a sequence you can actually run.
Does The AI Know Which Domains Are Yours?
It does. Asked from memory alone, the best model named related domains at 82% precision (Mashood and Nabeel, an arXiv preprint). Four out of five.
Precision here means: of the domains the model claimed belonged to the brand, how many actually did. And domains it failed to mention aren’t in that number at all.
The measurement is narrower than it sounds. The model was told to list exactly 50 domains for a brand, and the score is how many of those 50 checked out, averaged across the 36 brands.
Memory-only means no search either — no registration lookup, nothing but what the model absorbed during training. Add it up and a fairly accurate list of your domains is already sitting inside the model — which, from where marketing sits, is the finding.
Listing Them And Vouching For Them Are Different Skills
The authors kept the memory-only condition and changed the question. Now it’s yes or no: does this domain belong to this brand? Macro F1 topped out at 0.37, across the same 36 brands.
Macro F1 rolls two things into one score — did you catch the real ones, and did you keep the wrong ones out. A perfect score is 1.0, so 0.37 sits a long way down.
Listing works. Vouching doesn’t. Same models, same brands, same memory-only setup, and only the question changed.
Reeling off names from the class list is easy. Standing at the reunion door and saying whether the person in front of you was in your year is a different job.
So where does the gap come from?
Give Them The Registration Record And The Yes/No Call Reaches 0.99
Add to what the model can consult, and the score on the same yes/no question climbs by up to 0.65 macro F1 points — from a ceiling of 0.37 to above 0.85.
Precision comes in at 0.99. That figure covers one thing: of the domains the model said yes to, how many really were the brand’s. From memory alone it already sat between 0.84 and 0.99, because a model that almost never says yes is right on the rare occasion it does.
What WHOIS moved is the misses. From memory alone, the models called 74% to 92% of genuinely brand-owned domains somebody else’s. With the lookup available that falls to between 20% and 33%.
WHOIS is the public lookup that returns who registered a domain — registrant name, registering organisation, registration date. Anyone can query it.
The evaluation ran four consultation conditions:
- Memory alone
- Web search
- WHOIS lookup
- Combinations of those
One outside source, and the numbers on the same question move that far. Since the question stays fixed while only the source the model may consult changes, the gain reads as what the outside lookup contributed.
Still, that 0.99 came out of an evaluation that handed the model WHOIS on purpose, and on an ordinary day nothing guarantees a model runs that lookup every time.
The authors go on to recommend how to slot large language models (LLMs) into brand-protection work.
And here’s where I think marketing has something to do. Whether an AI can vouch for who owns your domain turns out not to rest on how much the model memorised. It rests on whether there’s a record outside that it can check against.
All you can do about the model is wait. The record you can touch.
How Far Does 0.99 Travel?
This is a preprint (arXiv:2606.20868), so no third party has been over the method or the numbers yet. The authors’ affiliation appears neither on the paper’s title page nor in the arXiv record, and there’s no venue listed either, so there’s no sizing the work up by whose lab it came out of.
The motivation is anti-phishing, and nothing here was designed as a visibility measurement for marketers.
The conditions are narrow in places:
- The brands are the 36 most targeted by phishing, which skews heavily famous
- The models evaluated are Gemini 2.5 Flash, Gemini 3.5 Flash, Claude Sonnet 4.5 and Claude Sonnet 4.6
- ChatGPT isn’t among them
Whichever model you end up asking when you try this yourself, it may well not be one of those four. And for a small or mid-sized brand, there’s no guarantee the same 82% shows up. The 36 names at the top of a phishing target list sit in a different category of how often the training data mentioned them.
Brands with a small public domain footprint came in between 0.32 and 0.60. Asked for 50 and short of real ones, the models fill the quota with plausible domains that don’t exist.
The authors do set out their own limits, and one of them lands right here: the evaluation drew 50 owned and 50 non-brand domains per brand, and they note it may not carry over to brands holding fewer domains. They also flag that models were held to two tool calls per question, and that a domain registered after the training cutoff tends to come back as “not owned”.
One more thing about WHOIS: registrant name and organisation can be hidden behind a privacy proxy, a service that puts the proxy provider’s name into the record instead of the registrant’s. Companies that hide them have their reasons. So “publish everything immediately” isn’t the conclusion.
What transfers to your own company is the checking sequence. The 0.99 stays with the conditions it was measured under.
Have The AI List Your Domains, Then See Whether The Record Backs Them Up
Put your internal domain register on one side and, on the other, the list a model gives back when asked which domains belong to the company. Laying those two next to each other surfaces both the ones missing from the register and the ones you don’t recognise.
If 82% is the ceiling, figure on close to one in five coming back wrong, and that’s the best case. The register stays the reference. The model’s list is a first draft.
Then take the domains that are actually yours and check, with IT, whether the registrant name and organisation are visible from outside. Whether to lift a privacy proxy is each company’s call. But knowing comes first.
Talk about visibility without knowing whether the model’s list is off, and you’re a piece short. I hadn’t noticed the piece was missing until I opened our own list.
The attribute side of what an AI believes about your company is in Let’s Look At Where The AI Actually Files Your Company. This one is the address book.
So the domain registration record is a marketing question now. Of all the things this year was going to put on my desk.
Sources
- Fathima Mashood, Mohamed Nabeel, “Can LLMs Reason About Brand Ownership? An Empirical Study of Domain Attribution Intelligence”, arXiv:2606.20868, 2026-06-18, https://arxiv.org/abs/2606.20868