Competitor web analytics: why every traffic number about a rival is a model, and the four things you can check for real
Third party traffic estimates are inferences, not measurements. Here is what Google itself says about outside tools, which competitor numbers are exact, and the one competitor surface almost nobody checks, with our own dated readings.
Published by AI Knows Us · Updated 22 October 2026
No tool measures a competitor's traffic. It models it. Analytics data is first party by design, so every outside figure about a rival is an estimate built from panels, clickstream and extrapolation, and the vendor chooses the method. That does not make the numbers useless, but it does change what you are allowed to conclude from them. Four competitor facts are exact rather than modelled, and they are the ones worth building a decision on.
Google puts this plainly in its own documentation for people buying search work. Its guidance on whether you need an SEO, read on 22 October 2026, says that third party tools "don't have access to Google's internal ranking data" and warns readers to "be wary of tools that claim to be 'acceptable' or 'approved'". The same page carries the sentence that settles every forecast argument: "No one can guarantee a #1 ranking on Google." If the engine will not hand its own ranking data to a tool, a tool cannot be reading your rival's traffic out of it either.
What an estimate is actually made of
A competitor analytics product has to infer a number it cannot observe. The usual ingredients are a browser or device panel, purchased or licensed clickstream, search volume data, and a model that scales a small observed sample up to a whole market. Each of those steps carries error, and the error is not evenly spread: small sites, business to business sites and sites with heavy logged in use are the hardest cases, because a consumer panel rarely contains the right people.
We could not read one vendor's own methodology to quote it here. On 22 October 2026 support.similarweb.com returned a connection error to our fetch, so we are not quoting a figure or a method from it. A failed fetch records what happened on a date. It is not evidence that the vendor's method is bad, and it is not a licence to describe a method we did not read.
The honest way to use an estimate is as an ordering, not a quantity. If one rival is modelled at roughly ten times another, the gap is probably real. If two rivals are within a few per cent, the difference is inside the model's own noise and you should treat them as equal.
The four competitor facts that are exact
| What you can check | Where it is exact | What it tells you |
|---|---|---|
| Who you overlap with in the paid auction | The Auction Insights report inside your own Google Ads account | Real impression share overlap for the auctions you both entered, not a model |
| What a rival publishes | Their own site, sitemap and pricing or product pages, read on a date | Their claims, their limits and their omissions, quotable with the URL |
| What a rival is cited for in AI answers | A recorded set of prompts and the citation URLs each answer returned | Whether an assistant uses their pages as a source when a buyer asks |
| Whether a rival is named in AI answers | The same readings, counted once per reading | A different finding from citation, never added to it |
The first of those is the one most teams skip. Auction Insights is free, it sits in an account you already own, and it reports the real outcome of auctions you both entered rather than an estimate of them. The reports and the settings that quietly change their numbers are set out in our guide to paid search analytics.
The competitor surface with no incumbent tool
When a buyer asks an assistant for a supplier, the assistant produces an answer with a set of cited sources and, often, a set of named companies. Both are measurable, and both are measurable about your rivals as easily as about you, because the answer is public. This is the part of competitor analytics that a web traffic estimator cannot see at all.
Our own US reading on 22 October 2026 gives the shape of it. Across 635 scored readings on ChatGPT and Gemini, 19,793 citation URLs were returned from 3,307 distinct hosts, and half of all citations came from just 51 of those hosts. The roster of named vendors was 24 names across 1,053 namings, counted once per reading in which a name appeared. The full host table and the question set are published in the citation study, and the naming side, with its denominator, is in the naming count.
Two warnings travel with those numbers. Citation and naming are two findings and must never be added, averaged or compared. And a reading is only comparable to another reading taken under the same prompt set, the same engine and the same date, because the same prompt does not return the same brands twice: run to run, we measured 53.0 per cent of brand and prompt pairs present in all three ChatGPT runs and 26.1 per cent on Gemini across 132 readings.
We sell the thing this page measures, so the disclosure belongs here rather than at the bottom. AI Knows Us reads five AI surfaces and reports whether a business is named and whether its own pages are cited, as two separate findings. Our own result on 22 October 2026 was named 0 and cited 0 across 635 readings, which is in our own published null result. Where a rival is stronger, say so plainly: Profound publishes SSO and SAML, SOC 2 and up to nine engines, and our rendered pricing page carries none of those words and five surfaces, not nine. Our cheapest plan that reads all five surfaces is Growth at 299 dollars a month in the US and the UK; the 149 dollar SEO Start plan reads 0 questions, 0 credits and 0 engines, so it is not an AI visibility plan.
A competitor read you can run this week
Open Auction Insights for your highest spend campaign and write down every domain that appears, with its impression share. Then open each of those sites and record what they publish about price, scope and proof, with the URL and the date beside each line, because a vendor fact belongs to the page it was read on and not to the brand. Then take the ten questions a buyer in your category would actually type, run them on two assistants, and record for each answer the sources it cited and the companies it named. Keep the raw answers. That file is worth more than any dashboard, because you can hand it to somebody else and they can reproduce it.
Finally, write the denominator next to every number you carry out of the exercise. A rate without the count it was taken over is not a measurement, and the metric definitions and their denominators are set out in our metric dictionary.
Questions buyers ask about competitor web analytics
Can any tool see my competitor's Google Analytics
No. Analytics access is granted per property by its owner. Every outside number about a rival's traffic is modelled from other inputs, and the vendor decides the model.
Are the estimates worth paying for at all
They are worth it for ordering and for direction of travel, and they are not worth it for arithmetic. Use them to tell a big rival from a small one and to spot a sharp change, not to calculate a market share to one decimal place.
Which competitor number should I put in a board pack
Put the exact ones in: auction overlap from your own account, what each rival publishes with the read date, and the citation and naming counts from a recorded prompt set with their denominators. Keep modelled traffic estimates in an appendix and label them as models.
Why do two tools report different traffic for the same site
Because they are different models over different samples, not two readings of one meter. When they disagree, neither is wrong in the way a thermometer can be wrong, and the gap is a measure of the uncertainty rather than of the site.
How do I measure a competitor inside AI answers
Fix a question set, record the engine and the date, run it, and count citations and namings separately against the number of readings. Our own method, its failures and the prompt set are published so that the result can be checked rather than believed.
The number nobody in your category is reporting
Most competitor reporting is a traffic estimate with no denominator and no date. The two cheapest honest numbers are sitting unused: the overlap report in your own ad account, and what an assistant says to a buyer who has never heard of either of you.