Skip to content

Results from our own websites, measured the way we measure clients

Every figure here is read live from our measurement database when you load the page. Only websites we own appear. A client’s numbers stay theirs.

What our measurements show

Being used as a source and being named are different things

An AI assistant can build its answer on a website without ever naming the business behind it. A tool that only counts names would miss that entirely.

ReferLabs, latest measurement on 2026-08-10: named in 0.7% and used as a source in 31% of 147 answers.

Most questions already have a website the assistants lean on

On almost every question we measured, one website was used as a source in more than half the answers. Who holds a topic has to be measured before anyone spends on pages.

37 of 38 questions, from 624 graded answers collected 15 to 29 August 2026.

We only call a change when it is bigger than normal wandering

Answers vary from day to day. When a change is too small to separate from that, we say it is too close to call and publish it anyway.

Every verdict we have reached is listed below.

Every measurement of our own websites

Read from the database at the moment you loaded this page. What each figure means is explained on the glossary and the methodology pages.

430

Answers read and graded

6

Measurements run

3

Websites

14

Changes shipped and logged

ReferLabs

Comparison publishing, Australia

Owned and operated by us; the same person runs both businesses. Revenue depends on being the cited source, which makes it the strictest test we have.

2 comparable measurements, judged on how often the site is used as a source.

Named in

0.7%

90% range 0 to 1.7%

Used as a source

31%

9 pages used

Answers sampled

147

Perplexity, AI Overviews, ChatGPT

Most-named rival, share of voice

4%

Xero

  1. 2026-08-08used as a source 29%, 87 answers, first measurement
  2. 2026-08-10used as a source 31%, 147 answersToo close to call (+1.8 points)

Comparing the 2 measurements since the way the answers were collected changed (our own engine calls → a third-party daily feed). 1 earlier measurement is excluded: a different way the answers were collected is a different test, so that first measurement on the new one is the baseline.

Grounded, collected daily by a third-party feed (retired 12 Aug 2026; its engine list included Google AI Overviews, which our own rig does not probe). 10 fixed questions, question set 5ed18f790e1857b7.

LuxurySaudi4U

Luxury travel, Saudi Arabia

Also ours: a second vertical from a standing start, to show the method is not category-bound.

One measurement so far, so there is nothing to compare it against yet.

Named in

0%

90% range 0 to 1.8%

Used as a source

Not measured

Web search was off on that run

Answers sampled

148

ChatGPT, Claude, Perplexity

Most-named rival, share of voice

6%

Banyan Tree AlUla

  1. 2026-07-21named 0%, 148 answers

Ungrounded run, direct engine calls. 10 fixed questions, question set d2375f127bf19719.

Income Lab (this business)

AI visibility, Australia

We measure ourselves on the same frozen questions we would set for a client.

One measurement so far, so there is nothing to compare it against yet.

Named in

0.7%

Range shown once every question has repeat answers

Used as a source

0%

0 pages used

Answers sampled

135

ChatGPT, Claude, Perplexity

Most-named rival, share of voice

2%

Profound

  1. 2026-08-18named 0.7%, 135 answers

Comparing the 1 measurement since the web-search setting changed (off → on). 1 earlier measurement is excluded: a different web-search setting is a different test, so that first measurement on the new one is the baseline.

Grounded (web search on), our own daily collection. 10 fixed questions, question set 6d084be71d512d4f.

How we measure, and the raw data

A client’s report is built by this same pipeline. The methodology page sets out the method in full, and the data page carries the tables behind every figure, with JSON and CSV.

A full report on one of our own websites

Corrections

  1. The 90% range beside a named rate was shown whatever the sample. It now appears only when every question had more than one graded answer on each assistant, which is the rule the method states.

  2. Our figures are re-scoped to completed searches. 166 of Claude’s 261 answers from 15 to 29 August 2026 report their search stopping early while still carrying a source list, so the answers behind those figures fell from 790 to 624.

  3. We withdrew the finding that these categories are open. Of 624 graded answers collected 15 to 29 August 2026, 14 named no business at all (2%). What it claimed and what we found

  4. A finding on this page said one of our websites was the most-read source in its category. Over 15 to 29 August 2026 it was second, after reddit.com.