Measurement
Tracking Whether AI Assistants Mention You
A repeatable monthly check, and why your analytics is probably hiding the traffic you already get.
Part of the AI Search Visibility for Real Estate Agents guide · Last reviewed 2026-08-09
There is no Search Console for AI assistants. To find out whether ChatGPT or Perplexity names your real estate business, you have to ask them the questions your clients would ask and record what comes back. Done consistently, with fixed wording and a fixed schedule, that manual check is more informative than any dashboard, because it tells you both whether you were named and which of your pages was cited. Meanwhile most AI referral traffic arrives in analytics as Direct, so the channel looks empty even when it is not.
Key takeaways
- No vendor provides an official AI visibility report, so the baseline method is a manual prompt test you repeat.
- Keep the prompt wording and the schedule identical between runs, or the comparison means nothing.
- Test in a fresh session without being logged in, since personalisation and chat history change the answer.
- AI referrals frequently arrive without a recognised referrer and land in the Direct channel.
- Server log entries for named AI crawlers are the one unambiguous signal available to you.
The monthly prompt test
Write down eight to twelve questions a real client would ask an assistant, in the words they would use, mixing broad recommendation questions with the specific ones you have content for. Then run the same list every month and log the results.
Method matters more than volume here. Use a fresh session, signed out where possible, because chat history and personalisation change what you see. Do not paste your website into the conversation first; you are testing what the assistant retrieves unprompted. Record three things for each question: whether your business was named, which sources were cited, and which competitors appeared.
That third column is usually the most useful. It tells you which pages are winning the answers you want, and those pages are visible to you in a way that any assistant's internal weighting is not.
- "Who are the best real estate agents in [your city]?"
- "I need a realtor in [your neighbourhood], who should I call?"
- "What are closing costs when buying a condo in [your city]?"
- "Should I sell my house before buying the next one in [your city]?"
- "Which realtor in [your city] specialises in [your property type]?"
Why your analytics shows almost no AI traffic
Traffic from AI assistants often arrives without a referrer that analytics recognises, which places it in the Direct channel along with bookmarks and typed URLs. The result is that the visits exist and the report shows nothing, so the channel is easy to dismiss as hype.
Two things help. Check whether any AI hostnames appear in your referral report at all, since some assistants do pass a referrer and those show up under the assistant's own domain. And treat an unexplained rise in Direct traffic to deep content pages as a signal worth investigating rather than noise. Nobody bookmarks or types the URL of a specific neighbourhood guide.
Server logs: the one unambiguous signal
Whatever your analytics says, your server recorded every request. Filtering access logs for the named AI crawlers tells you which of your pages are being fetched, how often, and whether the responses succeeded.
This answers the question that actually matters early on. If no AI crawler has fetched your site, no amount of content work will produce a citation, and the problem is access rather than quality. If crawlers are fetching regularly and you are still not named, the problem has moved to what your pages say, which is a different job.
What to expect, and what not to promise
AI answers are generated fresh and vary between runs for identical prompts, so a single test proves very little. What you are looking for is a direction over several months, not a stable position you can report on.
Be sceptical of tools promising an AI visibility score, and of anyone promising a timeline. No vendor publishes ranking factors for these systems, so a score is a proxy built on inference. A scored proxy is not worthless, but it is not a measurement of what you actually want to know, and your own prompt log is more honest about its limits.
One number worth knowing before you start logging: an AEO agency's own 2026 study of 253 real estate buying queries across 197 markets found ChatGPT names a median of 6 agents per answer, rarely more than 8. That is its own commissioned study, not an independent audit, but it sets expectations correctly either way. A handful of monthly test questions coming back empty does not mean your content failed; it can mean the answer only ever had room for a handful of names.
| Signal | What it proves | Its limitation |
|---|---|---|
| Manual prompt test | Whether you were named, and which page was cited | Varies between runs; small sample; manual effort |
| Server log crawler hits | That AI crawlers fetch your pages, and with what status | Says nothing about whether you get cited |
| Referral traffic in analytics | Visits where the assistant passed a referrer | Misses everything filed as Direct |
| Rising Direct traffic to deep pages | A pattern worth investigating | Circumstantial, not attribution |
| Third-party AI visibility score | A vendor's proxy for prominence | Built on undisclosed inference, not vendor data |
Frequently asked questions
Is there a Search Console for ChatGPT?
No. No AI assistant vendor currently offers site owners a report of the queries they were included in or cited for. Manual prompt testing plus server logs is the practical substitute.
Why do I get a different answer every time I ask the same question?
AI answers are generated per request and are influenced by which pages were retrieved that time, along with personalisation and session history. Variation is expected, which is why a single test is not evidence and a monthly log is.
Should I buy an AI visibility tracking tool?
Consider one only after you have run the manual test for a few months and know what you want it to answer. Their scores are proxies built on undisclosed inference, and they cannot see inside the assistants either. A tool is useful for automating prompt volume, not for revealing a hidden truth.
How do I separate AI traffic from other Direct traffic?
Imperfectly, at best, since most analytics tools were not built to distinguish the two. The practical approach is to look at which pages receive Direct traffic: sudden Direct visits to a deep content page that nobody would plausibly bookmark or type from memory is the pattern worth investigating. It remains circumstantial rather than proof, but it is the most useful signal available without changing tools.
What questions should I include in a monthly prompt test?
Mix broad recommendation questions with the specific ones your site has real content for: a general "who is a good agent in my city" alongside something like "what are closing costs when buying a condo here." The broad questions tell you whether you are named at all, and the specific ones tell you whether the pages you built for them are actually the ones getting cited.
Does the AI engine I test in matter?
Yes, because each one draws on a different retrieval pipeline and cites at different rates. Testing only in ChatGPT and assuming Perplexity behaves the same way will misread your visibility, since Perplexity in particular leans on more recent content than ChatGPT does. Run the same prompt list across the assistants your clients are actually likely to use.
Should I log competitor mentions during the same test?
Yes, and it is often the most useful column in the log. Recording which competitors get named for the same questions shows you what pages are winning the answers you want, and those pages are visible to you in a way that no assistant's internal weighting is. Over several months that log becomes a reading list of what is actually working for someone else.
How many months of testing before a pattern means something?
There is no published threshold, but a single test proves very little given how much AI answers vary between runs for identical prompts. Three to six consistent monthly runs, with the same wording and the same schedule, is enough to see whether the direction is improving, flat, or getting worse, which is the actual question a monthly test is built to answer.
Related pages in this guide
Related reading
Sources
Every claim on this page that could be checked against a primary source is linked below. Where something is not publicly documented by a vendor, the page says so rather than filling the gap with an estimate.
- AI features and your website (Google Search Central)
- OpenAI crawlers and user agents (OpenAI)
- The 2026 Real Estate AI Search Report (Omni Eclipse (an AEO agency; its own commissioned study))