Organic search traffic is falling at some sites and rising at others. The size of the fall depends almost entirely on which month you start counting from. We measured ten UK agencies over 24 months. Seven of the nine with a full baseline carry more organic traffic than they did in August 2024. Seven of the ten sit below their own peak. Both sentences describe the same table.
Neither is the headline you have been shown. What follows is the table, then the reason no third-party traffic estimate can settle this question for your own site.
Where does the sample come from?
The ten UK agencies holding the first page of Google for “generative engine optimisation agency”. They were selected for a different article, on a different question, before this one was asked. That matters more than it sounds. A sample chosen to answer “is organic traffic falling” can be picked to produce either answer.
Method. Ahrefs site-explorer-metrics-history, monthly, mode=subdomains, August 2024 to August 2026, pulled on 10 August 2026. One endpoint, one date, ten domains.
Two limits, stated because they bound everything below. The figures are worldwide rather than UK. Passing a country filter to this endpoint returned an empty error on our subscription, so we report the worldwide series rather than substitute a different source. And Ahrefs organic traffic is an estimate built from rankings, search volume and a modelled click-through rate. It is not a count of visits. Hold that thought, because the last section is about it.
The table
| Agency | Aug 2024 | Aug 2026 | Two-year change | Peak month | From peak |
|---|---|---|---|---|---|
| Havas Market | 2 | 944 | +47,100% | Aug 2026 | at peak |
| Push Group | 691 | 4,017 | +481% | Aug 2026 | at peak |
| Team Lewis | 7,201 | 24,525 | +241% | Mar 2026 | -30% |
| Reboot Online | 3,602 | 8,355 | +132% | May 2026 | -41% |
| IDHL | 1,313 | 2,342 | +78% | Sep 2025 | -26% |
| Impression | 28,034 | 30,181 | +8% | May 2025 | -37% |
| MarGen | 0 | 31 | from zero | Aug 2026 | at peak |
| Quirky Digital | 4,851 | 3,237 | -33% | May 2025 | -88% |
| Catalyst | 3,127 | 1,014 | -68% | Apr 2025 | -71% |
| Tilio | no data | 20 | n/a | Jul 2026 | -26% |
Tilio’s site carries no measurable organic traffic before February 2026. Its figures are too small to read as a trend, so it is in the table for completeness rather than for evidence.
What does the table actually say?
Most of them are up. Add the nine domains with an August 2024 baseline together and the sample went from 48,821 estimated monthly organic visits to 74,646. That is 53% more traffic than two years ago, across a set of businesses selling the fix for AI-driven traffic loss.
Most of them are also below their peak. Seven of the ten. Three are sitting at their highest recorded month right now.
Both readings come from the same twenty-five numbers. Nothing has been selected. The two-year column and the from-peak column disagree because they measure different things, not because one of them is dishonest.
The one entry that carries the collapse narrative properly is Quirky Digital: 26,304 in May 2025 down to 3,237 now, a fall of 88%. It is also the agency currently ranking second organically on the money term. Winning the query and losing the traffic are not the same event.
Where does “down 87%” come from?
Peak-picking, almost always.
Take two productised agencies from our earlier research. ManyPixels peaked at 115,160 estimated monthly organic visits in October 2024. FATJOE peaked at 48,052 in February 2026. Measure each from its own peak to today and you get two large declines. Measure from any other month and you get a smaller one. Neither choice is wrong. Only one of them gets written down.
Now the awkward part, which is a correction to our own numbers.
Our internal research file, pulled from the same endpoint with the same parameters on the same morning, recorded ManyPixels at 15,213 for August 2026 and FATJOE at 13,491. Re-run this afternoon, the same call returns 21,516 and 21,032. Every earlier month matches to the unit: 115,160 for October 2024, 48,052 for February 2026, identical in both pulls.
So the corrected figures are these. ManyPixels is down 81% from its October 2024 peak, not 87%. FATJOE is down 56% from its February 2026 peak, not 72%. Both remain steep declines. Both were overstated by measuring to a month the index had not finished settling.
Why did the same query return four different numbers?
Because a traffic estimate is a set of modelling choices, not a reading off a meter. Here is one domain, one month, one endpoint, on one day.
| The call | ManyPixels, Aug 2026 |
|---|---|
mode=subdomains, traffic_mode=static (the defaults) |
21,516 |
mode=subdomains, traffic_mode=adaptive |
19,086 |
mode=domain |
0 |
| The defaults, queried earlier the same day | 15,213 |
The zero is not a bug. The site’s traffic sits on the www subdomain, so restricting the call to the bare domain finds nothing. Anyone who queries a domain without checking that setting will report a healthy business as dead.
The gap between 21,516 and 15,213 is the one worth sitting with. Same parameters, same day, a 41% difference. We cannot see inside the index, so we are not going to guess at the mechanism. What we can say is what it costs you: a percentage change calculated against the newest month in one of these series is measuring the estimator as much as the website.
None of this makes the tools useless. It makes them the wrong instrument for one specific job, which is answering “are my clicks falling”.
So what should you measure instead?
Your own Search Console. It is the only source that counts your actual clicks. It is also free.
Three readings, in this order.
- Impressions against clicks, sixteen months, whole property. If impressions hold and clicks fall, you are being shown and not chosen. That is the shape everyone means when they say AI is taking the traffic. If both fall together, you have a rankings problem instead. That has nothing to do with AI.
- The same split for your ten highest-value pages. Site-wide averages hide the thing you care about. A brand term collapsing looks identical to a buying term collapsing until you separate them.
- Average position against click-through rate, by query, year on year. Holding position while CTR falls is the single cleanest signal that something above you in the result is answering the question. It also tells you which queries, which is what you act on.
That takes about twenty minutes. It produces a number about your business rather than a number about a market.
What this does not settle
Ten agencies are not an industry. They are ten small-to-mid UK marketing businesses whose traffic moves for reasons that include hiring, publishing, migrations and losing a single high-volume page. Do not read this as a UK B2B benchmark.
Worldwide figures for UK companies. Team Lewis operates globally. Quirky Digital does not. The same column means different things across the rows.
An estimate cannot separate causes. Nothing here shows whether a decline was caused by AI Overviews, an algorithm update, a competitor or a redesign. It shows the shape, not the reason.
Rising traffic is not good news on its own. A site can gain visits on queries no buyer types. That is most of what happened to the productised agencies before the fall.
The part that does hold
Whatever the direction, the click is a worse proxy for demand than it was. A buyer who reads the answer above the results and never scrolls counts as an impression and never as a click. Demand itself is the separate question, measured separately: UK search demand for buying SEO sits at a four-year high. Measuring that requires knowing whether you are inside the answer, which is a different question with a different method. Start with what AI visibility means and how to measure it. Then run the afternoon check if you would rather do it yourself today than buy a dashboard. If you are shopping for the dashboard, we compared the tools and their prices.
Data in this article: Ahrefs site-explorer-metrics-history, mode=subdomains, monthly, August 2024 to August 2026, all pulled 10 August 2026. Worldwide, because the country parameter returned an empty error on this subscription for this endpoint. ManyPixels and FATJOE figures re-pulled the same day and corrected against our earlier research, as set out above.