86% of an AI search citation share vanished in four days. Itera publishes its August 2026 LLMO/AIO trend report, free of charge

Itera, Inc. (head office: Shinjuku-ku, Tokyo; Representative Director: Takayuki Muto), which operates LLMOcheki, the SaaS product for measuring citation in AI search, published its monthly research report — the August 2026 LLMO/AIO trend report — on 8 September 2026.

It gathers what happened in LLMO through August 2026 into a single document: the real effect of the Google spam update, the unannounced change of underlying model, and the shift in citation structure, each set against measured data.

Why we publish it

Through 2026, AI Overview has become a normal sight in Japanese-language search as well, and “zero click” — users finishing their decision on the AI’s answer alone — has advanced quickly.

Meanwhile, companies have no established means of grasping how they appear in AI search. Google Analytics 4 measures only the clicks that arrive via AI; whether you were mentioned or recommended inside the body of an AI answer, or replaced by a competitor, cannot be measured.

In August 2026, a logging error that Google itself acknowledged occurred in the generative AI performance report in Search Console, and the Japanese and English versions of the official documentation on that fault were found to describe it differently.

Against this backdrop — the surfaces being built out rapidly while the means of measuring them fail to keep up — we organise the state of the market monthly and publish it in a form practitioners can act on.

Report summary

1. A month when the volume of discussion and the volume of actual damage were inverted

For the “August 2026 spam update” rolled out from 18 to 21 August 2026, our domestic fixed-point observation put the peak of volatility no higher than the threshold for a major shift. Over the same period, the citation share of one platform in ChatGPT search fell from 3.83% to 0.52% — a drop of about 86%. Ranking moved by a few percent; the change in citation structure was two orders of magnitude larger.

2. Search Console cannot detect this change

Increases and decreases in citation on AI search engines fall outside the tools Google provides. On top of that, the generative AI performance report in Search Console itself malfunctioned from 13 to 17 August 2026, so the “before the update” data used as a baseline was contaminated. It was a month that exposed the risk of measuring visibility through a single tool.

3. The underlying model changes without notice

On 11 August 2026 it was reported that Google had switched the model behind AI Overviews and AI Mode without any official announcement. There is a third-party report that a past model update replaced about 42% of the domains previously cited. Because the timing of a model change cannot be predicted, the only practical response is to have continuous measurement in place that detects change immediately.

4. AI Overview appears in 86.5% of Japanese-language searches (our own measurement)

In our measured study of 897 queries across 30 industry genres, the average appearance rate of AI Overview in Japanese-language Google search reached 86.5%. When AI Overview is shown, the click-through rate for the top-ranked result falls from an estimated 28.4% to 15.3% — a drop of about 46%. Traffic falls even where ranking is held, and this structure holds across a wide range of industries.

5. Urgency differs by industry by more than 30 points

The same study found the highest AI Overview appearance rate in professional services, SaaS and home renovation at 96.7%, and the lowest in hair salons at 66.7% — a gap of more than 30 points between industries. Meanwhile another study found that of 1,000 tax accountant websites, only 5.8% were judged ready for AI search: there are fields where AI answers appear most often and almost no business has responded.

What the report covers (18 items)

• A timeline of AI search in August 2026 (five weeks, 18 topics)
• The real effect of the August 2026 spam update, and a check on the “did AI-written articles fall?” argument
• Five things that happened at once in Google search in August, and how to tell them apart
• How the unannounced change of underlying model (the Gemini 3 and GPT-5.5 generation) affects citation
• Why “citation share” studies reach opposite conclusions depending on who runs them
• What determines citation probability (brand mentions 0.664 / backlinks 0.218)
• A comparison of citation logic across the eight major AI engines
• Separating “fetched” from “cited”, and “mentioned” from “cited”
• Data from Japan and abroad showing the shift in traffic structure
• The IAB’s “4 Ps of AI visibility”, and the data quality standard needed for budget decisions
• Entrants to the domestic LLMO market and the concentration of capital overseas
• Urgency by industry
• Five working principles drawn from August, and this month’s action checklist

Read the report (in Japanese) https://llmocheki.com/blog/news/2608-llmotorendreport

Comment from our CEO

“What was talked about most in August was the spam update, but line the measurements up and ranking moved only a few percent. Over the same period, 86% of an AI search citation share disappeared. The volume of discussion and the volume of actual damage were completely inverted.

I have worked in search and measurement for twelve years, across SEO and MEO, and I have never known a moment where it was this hard to say what your current position even is. Google’s own report can break, and the correction can be communicated differently depending on the language. Measuring visibility through a single tool is, I think, no longer realistic.

What August puts to us is a question: do you hold measurement of your own that can tell these things apart? I hope this report helps in building that.”

About LLMOcheki

LLMOcheki is a SaaS product that quantitatively measures how much a brand is mentioned and cited inside AI answers, across eight AI engines: ChatGPT, Gemini, Claude, Perplexity, Grok, Copilot, Google AI Overview and Google AI Mode.

n-shot statistical measurement: because a generative AI’s answer varies from run to run even for the same question, appearance rates are calculated from a sufficient number of runs with a Wilson 95% confidence interval.

Aggregation by engine: the overlap in cited URLs between AI Overview and AI Mode has been reported at 13.7%, so they cannot be judged combined and are recorded separately.

Separating mention from citation: the rate at which AI Overviews attach a link to a mention has been reported at 10.7%, so tracking links alone fails to capture actual exposure; the two are managed as separate metrics.

Daily API measurement with weekly real-UI measurement: two layers — fixed prompts, regular repetition, and a Japanese domestic environment — detect the effect of model changes and ranking shifts early.

LLMOcheki https://llmocheki.com/

Free LLMO audit https://llmocheki.com/llmo