We measured whether AI answers cite us. Here is the baseline

2026-10-02 · data from public, reproducible sources · by One IP

Measured 2026-10-08 with a grounded search API (Tavily), 5 buyer questions × 3 runs each.

Most "AI visibility" content tells you GEO matters. Almost none of it publishes its own numbers first. This is ours, before any of it is good news.

What we measured, and how

We took the questions a buyer actually asks before they hit "pay", asked them to a retrieval-grounded answer engine, and recorded what the answers actually cited.

• 5 questions, 3 runs each = 15 measurements
• Measurement success rate: 15/15 (no failed probes)
• Script: open, one command, source on the site

We are not going to report a score without telling you whether the instrument worked. A previous version of this check silently reported "0 citations" while every probe was failing with HTTP 429. Zero citations and a broken instrument look identical unless you measure the success rate separately. So we do.

The baseline: 2 of 15

2 of 15 measurements named us as a source. Zero of 15 mentioned us in the answer text itself.

We are in some retrieval results. We are in no answers.

Who gets cited instead

The clearest signal in the data is not about us — it is about what an answer is built from.

For "how to check if my IP is blocked by ChatGPT", the answers cited community.openai.com, reddit.com, plus two direct competitors of ours (cleanip.io and ipok.io — both do the same job as our AI-reachability check).

For "best IP reputation tool", 7 of 10 cited sources were other tool sites: easydmarc.com, apivoid.com, ipqualityscore.com, ipfighter.com, spamhaus.org, cloudns.net.

For "cleanest VPS IPs for a mail server", 8 of 10 were forums and comparison posts — lowendtalk.com twice, forum.cloudron.io, reddit.com twice, mailflowauthority.com.

The pattern: answers are assembled from whatever the web already argues about in public. Forums for opinion questions, tool pages for "what tool does X", and the competitors that showed up are the ones whose pages are shaped like an answer.

The one thing that works, and why

The only question where we appear at all is "check if my VPS IP is blocked by ChatGPT before buying" — and we hold rank #1 in 2 of 3 runs.

Compare it with the question we never appear on: "How can I check if a server IP is blocked by ChatGPT before I buy it?"

These are nearly the same question. The difference is not the topic. It is that the page we have for it is titled How to check whether your VPS IP is blocked by ChatGPT and structured as five concrete steps — and it matches the shape of the question.

Meanwhile the competing pages that beat us on the generic phrasing are tool pages that promise one action: paste your IP, get an answer.

So the rule we are taking from this is not "write more content". It is: make the page shaped like the question. A step-by-step article wins when the buyer asks "how do I"; a single-action tool wins when the buyer asks "check mine".

What we are changing, and what we are not

Changing: 1. Our AI-reachability pages are being rebuilt around one action, not a report. The evidence above says the generic question wants a tool, not an article. 2. We are testing whether our headings match the words buyers actually type. The winning question above is nearly verbatim our article title — that is probably not a coincidence.

Not changing (yet): 1. We are not chasing the brand-name gap we found. Searching our own name returns a different company (pure-ip.com, a voice provider) ahead of us. That is a real problem, but it is a naming problem, not a content problem, and we are not renaming. 2. We are not publishing a weekly score. One data point is a baseline, not a trend. We will report again when something has actually changed, not on a schedule.

The honest limit of this baseline

• 5 questions is not a market. It is a sample of what we guessed buyers ask.

The real set is larger and we have not found it yet.

• One engine is not "AI search." Tavily is grounded retrieval; ChatGPT, Perplexity

and Google AI Mode each behave differently, and we have not measured them.

• Rank #1 in a source list is not a citation. The answer text did not name us

in any of the 15 runs. Getting into the retrieval set is step one of about five.

• We do not know if this is good or bad for a site our size. We are publishing the

number precisely because we could not find anyone else publishing theirs.

*This is part of the work behind PureIP — free IP, network and AI-platform checks, no signup. The measurement script is one command; the baseline above is reproducible from it.*