Tools that answer a question are losing their searches. Tools that do a job are gaining them.
Seven years of monthly US searches for 5,716 online tools, sorted by what they give back. Since the year before ChatGPT, the median calculator, unit converter or code formatter lost 28% of its searches. The median file converter, timer or game gained 30%. The fall started slowly and is fastest now.
The median tool of each kind, against its year before ChatGPT
Each point is a tool's last twelve months of US searches against its twelve months to November 2022, which is 100. The line is the median across 4,344 tools that answer and 1,372 that do.
- Tools that do: convert a file, edit an image, run a timer, host a game
- Tools that answer: calculators, unit and currency converters, code formatters, word tools
Online tools come in two kinds, if you sort them by what they hand back. Some answer a question: what is 5 feet 9 in centimeters, what is the derivative of this, what does this JSON look like indented. Everything they return could be typed out as a reply. Others do a job: they turn a HEIC photo into a JPG, cut an MP3, count down a timer, deal a hand of cards. What they return is a file, a sound, a clock or a game.
Since the twelve months before ChatGPT was released, those two kinds have gone opposite ways. The median tool that answers lost 28% of its US searches. The median tool that does something gained 30%. Over the same years, the searches in the whole pool rose 17%.
Every kind of tool, sorted by what it gives back
The two groups are made of 14 smaller kinds, sorted by the words in each search. All 6 kinds of tool that do something grew. 7 of the 8 kinds that answer shrank, math and science most of all, at −48%.
Median change in searches since the year before ChatGPT, by kind of tool
The dot is the median tool of the kind; the bar is its 95% bootstrap interval.
- Tools that answer
- Tools that do
Table view
| Kind of tool | Gives back | Tools | Median change | 95% interval | Share that fell |
|---|---|---|---|---|---|
| Math and science | an answer | 957 | −48% | −52% to −44% | 75% |
| Unit conversion | an answer | 800 | −23% | −26% to −21% | 76% |
| Currency conversion | an answer | 111 | +35% | +25% to +45% | 22% |
| Code and data formatting | an answer | 414 | −29% | −35% to −24% | 73% |
| Words and writing | an answer | 126 | −13% | −22% to −6% | 64% |
| Lookups and reference | an answer | 56 | −7% | −22% to +4% | 57% |
| Everyday sums | an answer | 261 | −16% | −21% to −10% | 62% |
| Other calculators | an answer | 1,619 | −33% | −36% to −29% | 68% |
| File conversion | a job done | 114 | +36% | +16% to +50% | 28% |
| Images, audio and video | a job done | 426 | +39% | +29% to +48% | 27% |
| Makers and viewers | a job done | 500 | +23% | +15% to +30% | 38% |
| Timers and clocks | a job done | 55 | +35% | +9% to +64% | 29% |
| Games and randomizers | a job done | 245 | +27% | +19% to +37% | 30% |
| Device and skill tests | a job done | 32 | +48% | +17% to +128% | 19% |
The fall in code formatting fits an earlier study here, which found that searches for developer utilities stopped growing after 2023 while every other kind of tool search rose.
The exception is currency conversion, up 35%. It is the one answer here that goes stale: an exchange rate changes by the minute, so it has to be looked up, not remembered. The answers that fell are the ones that never change, the size of an inch, the rules of calculus, the shape of valid JSON.
Not a handful of tools, most of them
A median can hide a lot. So here is every tool, one dot for each ten, placed by how much its searches changed. The tools that answer sit mostly left of zero, the tools that do mostly right. 69% of the tools that answer lost searches (95% interval 68% to 71%), against 32% of the tools that do.
5,716 tools, by how much their searches changed
Change from the twelve months to November 2022 to the latest twelve. One dot is ten tools, rounded; the end columns hold everything past them.
- Tools that answer, above the line
- Tools that do, below
The far left of the chart is lopsided: 417 tools that answer, about one in 10, lost more than 80% of their searches, against 6 tools that do. Leave every tool past 80% down or fivefold up out of both groups and the medians barely move: −24% and +27%.
A Mann-Whitney test of the two sets of changes puts the chance of a gap this size between groups this large, if the two kinds were really the same, below one in 10173. The median tool that does something now gets 1.82 times the searches of the median tool that answers, relative to where each stood before ChatGPT.
It did not happen when ChatGPT arrived. It is happening now.
The timing is less tidy than the headline. The math and science tools had their own boom first: remote school sent them to 143 in the spring of 2021, and they had already fallen back to 100 by the time ChatGPT came out. They kept falling, to 52 today.
The math tools, the other tools that answer, and the tools that do
The same measure as the first chart. The lines mark November 2022 and January 2025.
- Tools that do
- Other tools that answer
- Math and science tools
The other tools that answer, the unit converters, calculators and code formatters, did not fall at first. For two full years after ChatGPT's release they held level at 100 while the tools that do kept climbing. Then, from early 2025, they started to drop, and they have lost 24% since. Before ChatGPT, the two kinds had grown almost in step: from 2019 to 2022 the median tool that does something gained just 3% more than the median tool that answers.
Since 2019, the median tool that does something has roughly doubled its searches. The median tool that answers gets 8% fewer than it did in 2019.
How this was measured
The data. Monthly US Google search volumes from January 2019 to July 2026, as Google Ads reports them, bought from DataForSEO in September 2026, for 7,689 searches that name an online tool and have a complete history.
The split. A tool answers when everything it gives back could be written as text in reply to a typed question: a number, a converted amount, a fact, reformatted code, a rewritten sentence. It does a job when it takes in or hands back a file, an image, a sound or a video, runs live against a clock, or is played. Each search is sorted by its words into one of 14 kinds, first match wins; the word lists are in the data file. 5,716 of the 7,689 searches fall into a kind; the other 1,973, which say too little to sort, are left out.
The change. For each tool: its searches in the latest twelve months (August 2025 to July 2026) over its searches in the twelve months before ChatGPT's release on 30 November 2022 (December 2021 to November 2022). The medians' 95% intervals come from 2,000 bootstrap resamples; the share that fell has a Wilson interval; the two groups are compared with a two-sided Mann-Whitney test (z = 28.0).
What this cannot tell you
- It does not prove ChatGPT did it. In the same years Google began answering questions at the top of its results, and people changed habits a search count cannot see. What the data does show is that whatever happened, happened to one kind of tool and not the other. A change in how Google counts searches would have moved both.
- Search volume is not use. A tool can lose searches and keep its visitors through bookmarks, or the other way round.
- Google Ads volumes are estimates, rounded into bands, with close variants merged. Medians across hundreds of tools are robust to that; a single tool's number is not.
- The pool is searches that name a tool job, collected in September 2026. It is not a random sample of all search.
The data
Every number on this page is in one file, the same file the page is rendered from. It holds aggregates only. Use it under CC BY 4.0: cite gizmobench.com and link to this page.
Download the data as JSON