Why this signal matters for Google
A benchmark is a measurement you took yourself and someone else can take again. It is not an opinion, and it is not a spec sheet comparison copied from manufacturer pages. You timed three hosting providers on the same page for two weeks. You counted how many clicks it takes to produce a quote in four invoicing apps. You publish the result, and you publish how you got it. That second half is what changes everything: it turns your number into proprietary data nobody can copy without redoing the work.
The people who score pages for Google are looking for experience, not a summary. Their scoring grid, the Quality Rater Guidelines, has them check in section 3.4 that the author has practised the topic rather than copied knowledge from somewhere else. Section 2.5.2 then sends them to find out who publishes the site and who signs the page. A test you ran yourself answers both questions at once: they see what was done, and they see who did it. An article that explains a topic could come from anybody. An article that measures something has to come from someone who did it.
Three pieces separate a test from proof: the method, the raw measurements and the limits. The method is the list of what you did, precise enough to repeat: which software version, which day, how many runs, on what machine. The raw measurements are the full file, not only the average: an average always hides the odd cases, and those are the ones a technical reader came for. The limits are the sentence almost nobody writes: "I only tested on one connection, results elsewhere may differ". Saying what your test does not prove makes you more credible, not less. On a YMYL topic (health, money, safety: Google calls those Your Money or Your Life), that sentence stops being optional, because a reader may decide something based on your number.
AI tools are hunting for numbers they can attribute to someone. Ask ChatGPT, Claude or Perplexity a question that calls for a figure: the answer leans on a source that carries a name, a date and a method. A page repeating "experts agree" gives them nothing to cite. A page saying "across 40 sites tested in March 2026 with this protocol, here is the median" becomes the source. The same thing happens with humans: other sites need sources worth citing, and a dated number with its method is far easier to cite than an opinion. The best-known example is Backlinko's study of Google ranking factors: by Backlinko's own count in 2025, more than 79,000 sites have linked to it. On the software vendor sites I audit, technical proof is usually the emptiest signal on an otherwise well-run site: the pages compare features, and none of them publishes a measurement taken in house.
The NeuroEEAT scoring grid (0-10)
Rate what a technical reader can reproduce from your page, without emailing you for the missing details.
| Score | Description | Concrete indicators |
|---|---|---|
| 0-2 | Nothing measured | No numbers of your own. The data comes from spec sheets or articles found elsewhere, often with no date and no link to the original. |
| 3-4 | A number with no instructions | You announce a result, sometimes with a nice chart, but nobody knows when, with what, or how many times. Impossible to reproduce, so impossible to argue with. |
| 5-6 | Method described | The protocol is written down: gear, versions, period, number of runs. A reader could rerun the test, but they only have your averages, not your measurements. |
| 7-8 | Reproducible | Full method, downloadable raw measurements, stated limits, visible test date. An attentive reader can rebuild your figures from your file. |
| 9-10 | Checked by others | All of the above, plus the script or procedure published as is, a rerun on a known date, and results from someone else who redid the measurement, including where they contradict you. |
How to audit it on your site
Take the page where you quote the most numbers and run it through these questions, one after the other.
- 01
List your numbers and trace where each came from
Take your ten most read pages and highlight every figure. For each one, answer in a sentence: measured by me, taken from a cited source, or came from nowhere. The third group is the one costing you.
- 02
Test reproducibility in five minutes
Hand the page to someone on your team who was not part of the test. Ask them to describe how they would rerun the measurement. If they have to ask you a single question, the method is missing something.
- 03
Check the raw measurements still exist
Find the original file from the test. If it is gone, or lives in a spreadsheet nobody can read anymore, your number is no longer verifiable, not even by you. A result you cannot rebuild goes back to being an opinion.
- 04
Look for the sentence about limits
Search the page for "limit", "only", "says nothing about". If no page states what the test does not prove, it reads as a sales claim, not a measurement. Add the most obvious limit: there is always one.
- 05
Date it and schedule the rerun
Does each test show the date of the measurement campaign, not just the article update date? A three year old software benchmark is wrong, not old. Put the rerun date in your calendar straight away.
How to activate or improve it
Start with a single measurement, taken this week, with the tools you already have. The rest builds up over time.
Produce one honest measurement
- Pick a question you already answer from memory, the "how long does it take" kind.
- Measure ten times instead of once, and keep all ten values, not only the best one.
- Write the protocol in five lines: what, with what, when, how many runs, how you counted.
- Add a limits sentence at the end of the section, saying what the test does not cover.
- Sign the test with your name and the date of the measurement, not just the article date.
Make the measurement reproducible
- Publish the raw measurement file as a download, in a format everyone can open.
- Give exact versions for everything you tested: a version number changes the results.
- Show the spread, not only the average: lowest value, highest value, and discarded runs with the reason.
- Declare what you stand to gain: partnership, affiliate link, free unit. A reader who finds out later throws out the whole test.
- Put the protocol on its own page, so you can link to it from every article that carries a number.
Let other people check it
- Publish the script or the exact procedure, so a reader can run the same measurement without contacting you.
- Rerun the test on a fixed schedule and keep the older versions online, so the trend stays visible.
- Send your results to the vendors you tested before publishing, and publish their reply even when it contradicts you.
- Invite a third party to redo the measurement and publish their number next to yours, with the gap explained.
- Correct in public when someone finds an error, keeping the old figure visible and dated.
5 common mistakes to avoid
Sector specifics
Limits before results
A risky topic for the reader: a number here can steer a care decision. A device or app test opens with what it does not measure, and the limit comes before the result, not in a footnote. Never compare clinical effectiveness from in house measurements: measure what you can actually measure, usability, waiting times, cost, and send the rest to official sources.
Separate the measurement from the promise
A measured past return is not a future return, and saying so is part of the proof. Test what genuinely measures: real fees over a year, transfer delay, support response time. Keep the exact period and the assumptions visible, because the same calculation over two periods produces two opposite answers.
Measure what you already do
No lab required. A roofer who records attic temperature before and after insulation across ten jobs has a benchmark, as long as the conditions are noted: season, exposure, roof type. Ten documented jobs beat a sales pitch, and nobody else in town has that file.
The comparison your prospects will redo
Your prospects will test it themselves, so assume your number gets checked. Publish the test data set and the script, state the versions, and accept losing on one criterion. A comparison you win on every line is not believed, even when it is true.
Signals to activate in parallel
One measurement rarely convinces on its own. These signals decide whether your benchmark is taken seriously:
FAQ
How to run a benchmark with no budget and no lab?
Measure what is already in front of you, and measure it properly. A stopwatch, a spreadsheet and ten repeats are enough to produce a number of your own. What makes a test credible is not the equipment, it is the written protocol, the kept measurements and the stated limits. A simple test anyone can redo beats a sophisticated one nobody can check.
How many runs make a test serious?
Enough that chance does not pick the winner, which in practice starts around ten repeats per case tested. The real guide is the spread: if your values sit close together, ten is fine; if they scatter, you need more, and above all you need to explain why they vary.
Should I publish the raw measurements or just the result?
Publish them. It is the cheapest move with the biggest return: it turns a claim into something checkable, and it protects you when a reader disputes the figure. An attached CSV is enough. Strip only what identifies a client, never what makes you look better.
What if my test favours a competitor?
Publish it as is, explaining which criterion they win on and why. It is the one moment a reader sees that your method outranks your interest, and it is what makes every other number on your site believable. A comparison you win on every point reads like an ad.
How often should a benchmark be rerun?
As soon as what you measured has changed: new version, new pricing, new plan. For software, once a year is the floor and six months is more honest. Show the campaign date next to the result and announce the rerun date: a reader forgives a dated number, never a stale number presented as current.
Case study
Before · a brand-new statistics page with no links
After · an email campaign to pages quoting outdated numbers
Source: Ahrefs (Joshua Hardwick), 2020, https://ahrefs.com/blog/link-building-case-study/ · The case is about a page of compiled, sourced figures, not an in-house measurement. It shows what happens to a data page when you offer it to the sites quoting outdated numbers: according to the article, the page reached first position for "SEO stats".