Primary research · September 2026
How AI Changed the Way People Search
What 261 model releases from 15 labs did to consumer behaviour, search demand and SEO.
- Every figure is measured from captured data, none are estimates.
- The full method, so you can re-run the audit on your own category.
- What to do about it, split by whether you run the company, the marketing or the site.
Nathan MzumaraAI Search & Digital Growth Strategist, twelve years in search
- 261
- Model releases
- 15
- Labs tracked
- 2.98
- Days between releases
- 68.0%
- Searches, no click
- 14.2x
- Cohort divergence
Free download
Get the full report
40 pages · PDF · 536 KB. Enter your details and it downloads straight away.
What is in this report
Every week somebody sends me a chart showing that AI has killed search, and every week somebody else sends me a chart showing search revenue at an all-time high. Both charts are real. I wanted to know which one describes the world the rest of us are actually living in.
So I did the boring thing. I took a workbook of every significant large language model released between 11 June 2018 and 3 September 2026, 261 of them from 15 labs, and read it against what has actually happened to search demand over the same window. Two datasets, one timeline. The AI industry narrates itself through capability and the search industry narrates itself through traffic, and neither conversation is honest about the other. A lab announces a context window, an agency announces a click-through rate, and nobody puts a date on the release that moved the click-through rate.
Search demand did not fall. 811 million searches a month are for the word chatgpt alone. What fell is the share of searches that end anywhere other than inside the answer, and it did not fall evenly. Queries an assistant can finish inside a tool you are already in fell 64.2% from their peak year. Physical, local and comparison queries fell 4.5%. Same index, same window, 14.2 times the damage. Which side of that line your demand sits on decides whether this report describes an inconvenience or an extinction.
A preview of the findings
3 of the 6 headline findings, in full. The rest, and the numbers behind them, are in the report.
2.98
A new model every three days, and the gap is still closing
Across 15 labs the mean gap between releases fell from 9.12 days in 2023 to 2.98 in 2026, a 3.1-fold compression. 95.0% of the dataset shipped in the last four years, so this is not a field with a long tail of history to reason from.
75.5%
Three quarters of every model ever announced is already dead
197 of 261 releases are superseded or retired, and only 62 are still the model their vendor is selling. Anything you build on has a median shelf life measured in weeks, not product cycles.
12x
The floor collapsed while the ceiling rose
The cheapest capable model fell from $0.60 to $0.05 per million output tokens. Frontier pricing went the other way, up 25%. Cheap intelligence is what put an assistant in front of everyone.
3 more findings in the full report
Download the free reportContents
29 sections across 40 pages.
- 01Why I built this5
- 02What is in the dataset6
- 03How I measured it7
- 04How far each source can be trusted8
- 05A new model now lands every 2.98 days9
- 0675.5% of every model ever announced is already dead10
- 07Only 62 models are on sale, and most are new11
- 08Models now hold 488x more in mind at once12
- 09Using a model got 12x cheaper, not dearer13
- 10The plain text-only model has stopped being made14
- 11110 models can now reason. In 2024 none could15
- 12Open-weight models peaked in 2024 and are now retreating16
- 13The labs have raised 12.4x more than they earn18
- 1458.3% of AI user numbers are not people choosing AI19
- 15ChatGPT's share of assistant traffic fell 23.7 points20
- 1668.0% of Google searches now end without a click21
- 17One search used to send one click. It now sends 0.5723
- 18Which searches AI took, and which it left untouched24
- 19Assistants cite Reddit 149x more than Stack Overflow26
- 20Assistants now send 771m visits a month, and it is growing28
- 21Google earns more from search while sending out fewer clicks29
- 22Nine predictions, with dates and falsifiers30
- 23What this changes in your job, by role32
- 24A 90-day plan you can actually run33
- 25What this report cannot tell you, and why34
- 26What I would test next35
- 27Glossary36
- 28The full data behind Parts II and III37
- 29Further reading and how to cite39
Free download
Get the full report
40 pages · PDF · 536 KB. Enter your details and it downloads straight away.