Two independent measurement systems, the same site, the same weeks — and a trend that turned out to be a lot less tidy than my first chart made it look.
In July, I published a piece on this platform about an odd pattern in my analytics: my small bilingual tech site, TechNovice, was logging more requests from AI systems fetching pages on a user’s behalf than it was logging actual human browsing sessions. The piece did better than almost anything else I’ve written here, and the framing — the biggest audience on my website never clicks — clearly touched a nerve for other publishers wondering the same thing about their own sites.
It also invited an obvious, fair challenge. A few readers raised it directly: I’d measured 89 days, ending right around the time I sat down to write. Was that window a fluke — a snapshot that happened to catch a temporary spike and mistake it for a trend?
That’s a legitimate question, and the only honest way to answer it is with more data, collected on purpose, over a longer stretch, and this time checked against a second, independent source instead of one dashboard alone. So that’s what I did.
Why One Data Source Wasn’t Enough
The original piece leaned entirely on one measurement system: my site’s bot-traffic log, which records every hit from a recognized crawler or user-triggered fetcher against every page. It’s a genuinely useful log. It’s also, on its own, a single instrument reading a phenomenon that’s hard to observe directly. User-agent strings can be wrong. Categorization can shift. A single busy week can look like a permanent trend if you stop measuring right after it.
So for this follow-up I tracked the same kind of window as before — 88 days this time, May 18 through August 13, 2026 — through two completely separate systems that have no particular reason to agree with each other unless something real is happening:
- The same bot-traffic log as before, isolating hits specifically from the “user-bot” categories — the fetchers tied to a live user asking a live question, as opposed to model-training crawlers.
- Google Search Console’s own dedicated report on the site’s performance inside AI Overviews, Google’s generative answer feature — a different company, measuring a completely different thing: impressions inside its own interface, not fetches against my server.
If both sources told a similar story, that would be much stronger evidence than either alone. If they disagreed, that would be interesting too, and more honest than quietly picking whichever number made the better headline.
They did both, depending on which weeks of the window you look at.
What the Bot Log Showed: Up, Then Down, Then Down Some More
Across the clean 88-day window, the site logged 68,495 hits from AI user-bots, almost entirely OpenAI’s user-triggered fetcher, with a small contribution from Perplexity’s equivalent. That’s a daily average of 778 hits, and 98.9% of them came back with a clean, successful response.
But an average hides the shape, and the shape is the actual finding.
The opening week (May 18–24) averaged 634 hits a day. Volume climbed steadily from there, cresting in the first week of July at 1,117 hits a day — 76% above the opening week — with a single-day high of 1,227 hits on July 6. That peak landed just four days before the data cutoff I used for the original article, which in hindsight means the July piece captured this pattern very close to its high point.
Then it fell. Not for a day or two — for the rest of the window. Every full week from early July onward logged a lower daily average than the week before it. By the last complete week I have (August 7–13), daily volume had dropped to 525 hits a day: 53% below the peak, and — this is the part that actually surprised me — 17% below where the whole window had even started back in May.
Read on its own, this signal doesn’t say “AI bot traffic is exploding.” It says AI bot fetch volume did a round trip, and landed lower than where I first started measuring it.
A Second Signal, Measured Independently, Tells a Different Story
Google’s AI Overview report, pulled for the identical window, shows a different shape for the second half.
Total impressions inside AI Overviews for the period: 171,036, averaging 1,944 a day. The opening week averaged 1,329 a day. Like the bot log, this metric rose through June and into early July, and it softened over almost exactly the same stretch in mid-to-late July — the low point was a single day of 1,161 impressions on July 24, in a week that averaged 1,378 a day. For about three straight weekly steps, two systems built by different companies, measuring different things, moved in the same direction at the same time.
I didn’t want to just eyeball that and call it a pattern, so I checked it: the 7-day rolling averages of the two series correlate at 0.77 across the first nine weeks of the window, through July 27. That’s a genuinely strong relationship for two metrics that share no infrastructure and no methodology.
import pandas as pd
bot_hits = pd.read_csv("bot-traffic.csv").groupby("date")["hits"].sum()
ai_overview = pd.read_csv("ai-overview-export.csv").set_index("date")["impressions"]
combined = pd.concat([bot_hits, ai_overview], axis=1, keys=["bot_hits", "ai_overview"]).dropna()
rolling = combined.rolling(7).mean().dropna()
print(rolling["bot_hits"].corr(rolling["ai_overview"]))
Then, in August, the two signals split. Where bot-log volume kept sliding, AI Overview impressions snapped back hard, up to 2,362 a day in the first week of August, and the final complete week of the window (August 7–13) still averaged 1,867 a day. That’s 41% above the opening week. Not a round trip. Still up. Run the same correlation on just that final stretch, from July 27 to August 13, and it flips to –0.45. Same site, same underlying weeks, two credible measurements, opposite signs.
So one legitimate, cross-checked AI-visibility metric ended the window below where it started. Another legitimate, cross-checked AI-visibility metric, measured on the same site over the same weeks, ended the window well above where it started. Both are real. Neither one is “the” number.
What Else Showed Up in Google’s Own Data
This was the first time I’d pulled this specific report, and it surfaced a few patterns worth sharing at the level of shape rather than specifics — for the same reason I gave in the original piece: some of this is the actual implementation, and I’d rather it stay that way.
- Concentration is extreme. Roughly 220 individual pages picked up at least one AI Overview impression during the window. Just ten of them account for 72.6% of all impressions. That’s a textbook power-law distribution, and it matches what I’d expect from citation behavior generally: a small number of pages do almost all of the work.
- It isn’t confined to the site’s main topic. About 69% of impressions landed on the eSIM and travel-tech content that’s the site’s core focus. But roughly 17% landed on an unrelated streaming and geoblocking guide, and about 12% landed on AI-tool explainers that have nothing to do with eSIMs at all. Whatever drives inclusion in AI Overviews doesn’t seem to care much about a site’s topical focus; it looks more sensitive to how a given page is structured than to what the rest of the site is about.
- The audience is genuinely international, and not mobile-first. Germany and the U.S. lead, at 20.8% and 16.8% of impressions respectively, followed by the U.K., Canada, and Australia, a mix that tracks the site’s bilingual DE/EN strategy reasonably well. Device split was close to even: 50.1% desktop versus 48.5% mobile. I’d have guessed AI-mediated search skewed more mobile than that. It doesn’t, at least not here.
Humans, Still Outnumbered, Now With a Precise Ratio
The original headline claim was directional: more AI-bot hits than human sessions, without an exact ratio. This time I have one, from the same clean window on both sides.
68,495 AI user-bot hits against 3,429 human sessions (3,339 unique visitors, 4,346 pageviews) works out to a 20-to-1 ratio.
The same caveat from the original piece still applies, and it’s worth repeating rather than quietly dropping: a hit is not a person, and a session can contain several page views while a single AI interaction can also generate several fetches. Twenty-to-one is not “twenty AI users for every human reader.” It’s twenty logged machine retrievals for every one logged human browsing session, a real and fairly extreme gap, just not the literal head-count comparison it might sound like on first read.
Three Guesses for the Divergence, and Why I’m Not Confident in Any of Them
I don’t have a clean explanation for why one signal kept falling while the other rebounded, and I’d rather say that plainly than invent a tidy story to fill the gap.
Three possibilities, none confirmed:
-
Google may simply be showing AI Overviews more often, independent of my site. If Google expanded the feature’s rollout or trigger rate for relevant queries during this window, impressions would rise site-wide for reasons that have nothing to do with anything I did.
-
OpenAI’s fetching behavior may be shifting toward caching. If ChatGPT is relying more on previously ingested content and re-fetching live pages less often for queries it has effectively already “seen,” raw fetch volume could fall even while the content’s actual influence on answers stays flat or keeps growing.
-
August is a genuine seasonal trough across parts of Europe. Lower query volume during peak vacation weeks could plausibly touch both metrics, just not identically, given how differently each one is generated.
Any of these could be true. More than one could be true at once. Telling them apart with confidence would need data I don’t have, and honestly, data no single publisher has access to. Anyone who tells you they know precisely why a crawler’s fetch rate moved in a given week is guessing, same as I am here.
For context on why this is worth tracking at all rather than dismissing as a curiosity: search interest in “AI Overviews” as a topic has more than doubled over the past year, and interest in “AI citations” specifically has been on a similar climb. This isn’t a niche question a handful of publishers are asking.
What This Actually Changes About How I’ll Report This
One clean lesson from this round: a single 89-day window, measured once, from one source, is a snapshot. Useful, but easy to mistake for a trend line if you stop watching right after a peak. I did exactly that in July, not through any dishonesty, just through timing. The extended window makes it obvious how much can change in the eleven weeks after you hit publish.
Going forward, I’m treating this the way I’d treat any metric I actually cared about getting right: rolling windows instead of fixed snapshots, at least two independent sources before I call anything a trend, and an explicit habit of reporting it when two credible signals disagree instead of quietly picking the one that tells the better story.
I’m also running a small, controlled test on a slice of my own content right now, designed to isolate one specific variable’s effect on this pattern. It won’t be conclusive for another six to eight weeks. When it is, that’s the next piece.
What I’m Still Not Publishing
Same policy as the original piece, for the same reasons. I’m not sharing which specific pages drive the concentration described above, the exact technical changes behind any of these numbers, conversion or revenue data, or the details of the test currently running. That’s the actual implementation, and publishing it wouldn’t make this analysis more useful to another publisher; it would just make my own site easier to copy.
What I think is genuinely useful to share is the shape of the pattern, the size of the gap between machine and human traffic, and, maybe most usefully this time, a concrete example of what happens when you refuse to trust a single dashboard: sometimes it protects you from overclaiming, and sometimes it hands you a genuine disagreement you can’t paper over.
Final Thought
Three months ago I wrote that my website’s biggest audience never clicks. That’s still true. Twenty-to-one isn’t a ratio that reverses in one quarter. What I’d add now is a second sentence: that audience isn’t one thing, measured one way, moving in one direction. It’s at least two different signals, built by two different companies, and for about nine weeks this summer they agreed with each other closely. Then they didn’t.
That disagreement is the actual finding here. Not the growth, not the gap versus human traffic, but the fact that this is genuinely harder to measure cleanly than a single viral chart makes it look, and that anyone reporting one confident number, mine included three months ago, is probably reporting a snapshot, not a law.