Is Keyword Search Volume Accurate? We Checked Against Ours

By Ugur Saritepe · September 30, 2026

Keyword search volume is accurate about topics and wrong about phrases. On 2026-09-30 I pulled three different keywords on my own account and got one figure back for all three: 32,987 monthly searches. One of the three is the head term the other two contain. Then I pulled the phrase my site actually earns impressions for and got 0.

Why I was looking at volume numbers at all

game-scout.app is my games site and my test bed. On 2026-09-11 I asked the question that started this:

How do we find comparison to run ?

The site publishes head-to-head comparison pages, and I wanted the routine that builds them to pick the next pair by itself. The first thing that fell over was the obvious source. Search Console only reports queries a page already gets impressions for, so every "X vs Y" query the property holds maps to a comparison we have already published. It is an excellent record of what you rank for and it cannot name a keyword you do not rank for yet.

That comment is still sitting in the picker script we shipped on 2026-09-11, because it was the whole reason the script needed a second demand source:

// Scope note, deliberate: this picker supplies SUPPLY, not demand. Search
// Console cannot discover a new pair (it only reports queries a page already
// gets impressions for, so every "vs" query it holds maps to a comparison we
// have already published).

So the pipeline started asking a keyword tool. Eighteen days later I have three months of its output to read, and this is what it says.

The same number came back for three different keywords

Six pulls on 2026-09-30, through the MCP server on my own account. Every one of these responses labelled itself confidence: "low", which is the tool being honest about what follows. Autocomplete presence is 1 or 0: is the phrase offered as a suggestion or not.

keyword                              autocomplete   bing monthly
resident evil 4                                 1         32,987
resident evil 4 vs resident evil 1              1         32,987
resident evil 4 vs revelations 2                1         32,987
discovered currently not indexed                1              0
crawled currently not indexed                   1              0
seo mcp for chatgpt                             0            449

The top three rows are the finding. resident evil 4 is a franchise head term. The two rows under it are specific comparison questions between named games, and one of them, Revelations 2 against Resident Evil 4, has no dedicated editorial page anywhere on the first page of Google. All three returned 32,987. That is not three measurements that happened to agree. It is one measurement, attached to every phrase that contains the head term.

A volume figure on a long-tail phrase is usually the volume of the topic the phrase belongs to. Pull the head term alongside your long-tail. If the two numbers match, you are reading the topic's number, and it tells you nothing about the phrase.

Google documents this upstream. Its Keyword Planner help page defines average monthly searches as the number of times people searched a keyword and its close variants, and says the volumes are rounded. Grouping is not a bug that one provider has; it is how the source data is built. What surprised me is how wide "close variants" turns out to be in practice: wide enough that two different game comparisons and their shared franchise name all resolve to one figure.

And 0 came back for the phrase my site earns most

Rows four and five are the opposite failure. discovered currently not indexed returned 0 monthly searches. Here is what my own Search Console says about that phrase over the 90 days to 2026-09-30, on traffictriage.com, grouped by query with a 250-row limit and 68 rows returned:

query                                   impressions   position
discovered currently not indexed                 52       85.0
discovered - currently not indexed               44       79.9
seo mcp for chatgpt                              10       73.2
why is my page not indexed                        7       82.0
soft 404                                          6       52.5
property total                                2,466       24.7

discovered currently not indexed is the single largest query row on the property. Counting every spelling and question form of it, the family runs to 25 rows and 132 impressions, which is 5.4% of everything the site was shown for in the window. The tool rated the head of that family at zero.

Two honest limits on that comparison. The volume figure is Bing's and the impressions are Google's, so a phrase really can be close to nothing on one engine and present on the other; that alone could explain part of a zero. And 132 impressions in 90 days is a small number in absolute terms, so this proves people type the phrase, not that the phrase is worth much. What it does rule out is reading a 0 as "do not write this". A zero means the provider has no row for that string.

The second signal failed in the opposite direction

The same response carries an autocomplete presence flag, and row six is where that one breaks. seo mcp for chatgpt returned presence 0, meaning the phrase is not offered as a suggestion. The same Search Console pull has it at 10 impressions at average position 73.2, the largest row in the property's eight-row family of assistant-connection queries.

So on the two phrases where I hold measured data, one signal said zero volume and the other said not-in-autocomplete, and both were wrong in opposite directions. Presence is also a weak discriminator on its own: it returned 1 for five of the six keywords above. The only control I have that came back 0 is a deliberately absurd pair, assassin's creed freedom cry vs assassin's creed odyssey, pulled on 2026-09-30. One control is not a calibration.

The command-line version was worse

There is a second path to the same Bing data, a keyword-research command that returns a hundred related terms instead of one figure. On 2026-09-29 a build run asked it for mechanic games, because the target page was a roundup of vehicle-repair simulators. The first rows it returned:

auto repair shops near me      mobile mechanic near me      repair
how to repair drywall          what is a pwa                how to repair chipped paint
how to repair zipper           how to repair screen door    automotive
system mechanic                car repair near me           mechanics

Not one row in a hundred was about a game. It gave back local car repair, home DIY, a Windows utility called System Mechanic, and what is a pwa. We had already found the same thing on 2026-09-12, when the command returned a topic cluster with no head-to-head rows in it for four different game pairs. A related-keyword list expands to the most-searched neighbourhood of your words, and if your niche is a small corner of a big word, you get the big word.

What actually decided the three pages

Three pages were built automatically on 2026-09-29 and 2026-09-30 while this data was coming back. In all three the volume figure did not decide anything. The search results page did.

Resident Evil Revelations against Revelations 2, 2026-09-30. The figure was 173,672 monthly, and the run wrote in its own pull request that it treated that as a topic-cluster sum and did not qualify on it. What qualified the page: 8 of 9 organic results were forum threads, one of them titled "Revelations 1 VS 2, Which One is Better?", and the only editorial result was a 2015 review that covered both games and declined to pick.
Resident Evil 4 against Revelations 2, 2026-09-30, the 32,987 pair. Qualified on a page one that was entirely forums, with a Reddit thread at position 1, plus a sweep of eight games outlets that found zero head-to-head pages.
A vehicle-repair games roundup, 2026-09-29. Both target phrasings showed autocomplete presence, page one held no dedicated roundup at all, and the top organic result was someone asking which of two repair games to buy. The run recorded that absolute volume was unverified and shipped anyway.

Read together, those three are the pattern. The number was noise in all three. The thing that told us a page was worth writing was that real people were asking the question in forum threads and nobody had answered it properly.

How to check this on your own keyword

Four steps, about five minutes, no paid tool. Stop at the first one that settles it.

Pull the head term next to your long-tail. Take your phrase, strip it back to the two or three words at its core, and get a volume figure for both. Identical figures mean you are reading one number for both, and the long-tail is invisible inside it.
Look up the phrase in your own Search Console. Group by query over 90 days. If a row exists, you have a measurement and you can stop estimating. Remember that Search Console names only some queries and withholds rare ones, so an absent phrase means not visible, never zero.
Read page one for the phrase. Forum threads, Q and A sites and a group post at the top mean people are asking and nobody has published a real answer. That is the strongest buy signal on this list, and it is the one the three builds above actually used.
Only then use the volume figure, and only comparatively. It can tell you topic A is bigger than topic B. It cannot tell you what one page will earn, and a 0 is not a verdict.

The one published comparison I could find that sets tool volume against measured impressions is Ahrefs' own help page, which reports its estimates as roughly accurate for about 60% of the keywords it studied against Search Console impressions, against 45% for Keyword Planner. It is a vendor grading itself and it names no keyword, but 60% is a fair headline for what a volume figure is: right more often than not, about the topic, and not something to plan a page on.

What we have not fixed

Two things in this pipeline are wrong and known, and both are still wrong at the time of writing.

The skill file that runs this still names the command we stopped trusting on 2026-09-12, in two places, and we have not corrected it. Three consecutive runs quietly used the other tool instead, which works until someone reads the file and believes it.

And one candidate pair has been skipped five runs in a row because of a length value in our catalogue we know is wrong, 520 hours for a game that is nothing like that long, and we have not fixed the row.

What to do with this

If you are choosing what to write next and a volume figure is what you are choosing on, you are choosing on the weakest of the three signals available to you. In order of how much it is worth:

Your own measured query rows. Nobody estimated them, and if you have any impressions at all you already know which phrases Google associates with your pages. The grouping method we settled on is how we turn those rows into topics.
Page one for the phrase. Forums at the top and no settled answer is the gap. This is what we used on the comparison pages and it is the only signal that has been right every time so far.
The volume figure, comparatively and with the head term pulled next to it. Treat a 0 as no data and treat a big round number as a topic.

This is the third measurement in this series where a Search Console number turned out to mean something narrower than it looked. The average position autopsy is the closest one, and the rest are in the AI-assistant SEO series. If what you actually want to know is whether Google can see and index the pages you already have, the free Triage Report runs eight checks from a URL and scores them on screen in about 30 seconds, no signup.

FAQ

Is keyword search volume accurate?

Not at the level of the phrase you typed. On 2026-09-30 the same tool returned 32,987 monthly searches for three different phrases on our own account, one of which is the head term of the other two. Google's own Keyword Planner documentation says the figure counts a keyword and its close variants and that volumes are rounded, so a number attached to a long-tail phrase is usually the topic's number. It is useful for saying one topic is bigger than another, and not for predicting what one page will get.

Why does a keyword tool show zero volume for something people clearly search?

Because a zero means the provider has no data for that string, not that nobody searches it. On 2026-09-30 the same tool returned 0 monthly for discovered currently not indexed, which is the single phrase our own Search Console recorded the most impressions for over the preceding 90 days: 52 impressions at average position 85. Zero volume and measured impressions can be true at the same time, and only one of the two is your own data.

What should I trust instead of search volume?

Three things, in this order. Your own Search Console query rows, because they are measured on your site and nobody estimated them. The search results page for the phrase, because it shows you what the reader is being offered and whether anyone has answered them properly. And the shape of the volume figure rather than its value, meaning whether the topic registers at all. The three autonomous builds we ran on 2026-09-29 and 09-30 all picked their topic on the results page, not on the number.

Does a paid keyword tool give better numbers?

We have not tested one, so we cannot tell you. Every figure in this post comes from one free Bing-backed signal, and no Ahrefs, Semrush or Keyword Planner pull was made to compare. Ahrefs publishes its own study saying its estimates were roughly accurate for about 60% of the keywords it studied against Search Console impressions, against 45% for Keyword Planner. That is the vendor grading its own product, and it is still the only published comparison to measured impressions we could find.

A volume number is a hint about a topic. Your own rows are a measurement of a phrase. Create a free account, connect Search Console, and the MCP server gives your assistant both in one session: the demand signals for a candidate keyword, with the confidence label on the response so you can see how much weight it carries, next to your own measured query rows for the same phrase. That is the comparison in this post, and it took me one message to run. The part that used to take an afternoon was noticing that the two disagreed.

Create an account

Claude running a scheduled TrafficTriage daily check: a Search Console report on a real site, summarized with clicks, impressions, and the one query cluster that lost rank

In your AI assistant

Ask your AI about your own site

TrafficTriage plugs into Claude, Cursor, and ChatGPT. Your assistant runs the 8 checks, reads your Search Console data, and can even check your site every morning. Free account, no API key, nothing to install.