Did we lose real people?

The totals fell. The real audience did not.

Users per day, so the 6-day window compares fairly. Bots is Singapore + China + Hong Kong. Real is United States + India. PostHog uniqExact, counted as a union, not a sum of columns.

Bots fell 81% on scalekit.com and 83% on docs. Real users per day on scalekit.com went up 30% over the same 10 weeks. Docs real users dipped in the 6 Jul week, then came back. Search clicks per day fell 3%. That is the whole story in one table.

6 Jul: PostHog and GA do not agree on docs

Week pair: 29 Jun 2026 to 5 Jul 2026 vs 6 Jul 2026 to 12 Jul 2026. If real demand died, both tools would fall.

docs.scalekit.com: PostHog users −25%. GA sessions +6%. That is not real demand dying. I think this is bots in PostHog, or mix.

Users and sessions fell together. Nobody was just staying longer.

A common explanation for a user drop is that fewer people visited but each stayed longer. That would show as users down and sessions flat or up. That is not this data.

Both fell by the same amount, so sessions per user did not change. Whatever left was visiting once and leaving, at the same rate before and after. That is the shape of a bot, not of a reader who lingers.

scalekit.com: one Singapore day, then gone

29 Jun week vs 6 Jul week. Almost all of scalekit.com's −11% is Singapore leaving. US and India went up.

Singapore / China / Hong Kong · 505 users on 29 Jun, then ~5–30/dayUS + India held

Daily cut agrees. 505 of those users sit on 29 Jun. Singapore / China / Hong Kong is where LLM agent bots run (proxies). I ignore this for demand.

docs.scalekit.com: China / Hong Kong fell 8–9 Jul. US + India did not.

Orange is Singapore / China / Hong Kong. Blue is US + India. Step is 8 Jul to 9 Jul.

Singapore / China / Hong Kong · 404 → 72 on 8–9 JulUS + India · 69 → 79 that same day

One segment fell (404 to 72). US + India that day went 69 to 79. GA sessions went up that week. Singapore / China / Hong Kong is LLM agent bots on proxies. I ignore this for demand.

docs.scalekit.com: GA sessions held. One page took the hit.

GA session_start, 29 Jun 2026 to 5 Jul 2026 vs 6 Jul 2026 to 12 Jul 2026. Site total 1,059 → 1,124 (+6%). This is real sessions, not PostHog bots. Top pages by week-1 share.

322 docs pages had a session that first week. The top 10% is 32 pages. /agentkit/quickstart/ is the one that dropped. Home held. Site total still went up.

scalekit.com Search Console impressions went 602k to 390k, Jun to Aug

Clicks held in July. Impressions did not. Ahrefs keyword count held. Positions 1-3 slipped. This is still the one that looks like a trend.

I already have the GSC queries file. It ends 5 Sep. That is the minus-2 cut. Ahrefs is not a mass deindex. Keywords in 1-3 went 214 to 181. Top 10 only went 423 to 411. I still will not name a reverse.

Google names more of my queries, not fewer. The lost impressions are the ones it will not name.

Query-level cut from gsc_site_queries.csv, 29 Jun–5 Jul vs 24–30 Aug. Both windows are 7 days. The query file is a subset of the property total, so I read its direction, not its size.

Read those four rows together. Named queries grew and carried more impressions. The property total still fell. So the loss sits in queries Google will not name, which is the anonymous long tail. I did not lose ranked search demand.

Still worth doing: positions 1-3 went 214 to 181 while 4-10 went 209 to 230. About 33 keywords slipped one tier. Total keywords held, 466 to 448, so nothing was deindexed. That is reversible with normal content and link work. It cost almost no clicks, so it is not urgent. I still will not name AI Overviews as the cause. Search Console gives web, image and news only. There is no generative column.

docs GA sessions fell 52% in the 10 Aug week. No page did it.

GA session_start, Monday weeks. This dip is the mirror of 6 Jul: the tool that missed the July cliff is the one that sees this. PostHog barely moved that week.

I checked whether one endpoint caused it. It did not.

A bad deploy or a lost ranking concentrates in a few URLs. This is 78% of all pages each losing 1 or 2 sessions. That shape is a crawler leaving, not readers leaving. Compare 6 Jul, where /agentkit/quickstart/ alone was 15% of the week and took the whole hit. These two dips have opposite shapes.

31 Aug week stops at 5 Sep. I am not calling it a trend.

Google lags about 2 days in GA and GSC. We pull until today minus 2. Today is 7 Sep, so 5 Sep is the right edge. Do not wait for 6 Sep. That day will look thin. I am not calling this a 7-day trend. The 6-day numbers are still below.

When did ai-train=no appear?

docs.scalekit.com has it. scalekit.com does not.

What it means: ask crawlers not to use docs to train models. Search indexing is still yes. Feeding docs into an answer at request time is still yes.

Should it be there? Common on docs sites. It is a preference, not a block. It does not stop Google Search. It was already there in April, so it did not cause the July drop.

New models shipped. I cannot tie them to this drop.

I said before that we had no LLM referral data. That was wrong. GA has an AI Assistant channel, and ga_site_referrals.csv names each assistant. Search Console still has no generative column, so this covers click-throughs from assistants, not AI answers that never got clicked.

scalekit.com sessions from LLM assistants, per week. Sources are chatgpt.com, gemini.google.com, claude.ai, perplexity.ai and copilot.com.

Flat at 22 to 46 all period. The highest week is 31 Aug, after every launch on this list. Ahrefs Web Analytics agrees on its own llm channel: 180 visitors in July, 192 in August.

Five tests would show it if new models took our readers. None do.

The broad version of this angle is dead. New models did not take our readers. Five independent tests say so, and the July cliff predates every launch by 5 weeks.

A narrow version survives. The traffic that left executed JavaScript, because GA counted it, and it arrived as Direct from proxy countries. That is agent browsing, not Googlebot, because Googlebot does not create GA sessions. New models do change which agents run. I cannot tie that to any single launch date, so I am not calling it.

I re-ran every number in the report against the raw sources

106 of 109 figures match exactly. PostHog on inclusive calendar dates, GA on session_start, Search Console on the page files, Ahrefs on the keyword history CSV.

The 706 matters only for method. uniqExact gives 706. count(DISTINCT) gives 707 because HogQL maps it to the approximate uniq(). The percentage does not change. The app.eu number was a 3-week slide, not one drop. Nothing here changes a single call on this board.

Four things to do. Three answer an open question. One is just watching.

Everything above is diagnosis. This is the whole action list that comes out of it.

Not on this list: brand search. The query scalekit lost 30% of its clicks while my ranking improved. That is a demand and awareness question, not a traffic bug. It needs a different owner and a different conversation.

Nothing more from this side.

Google lags about 2 days. Files already stop at 5 Sep 2026. That is the minus-2 cut.