How Does AI Interpret the Word "AIO"?

Netsujo Inc. · Published July 5, 2026 · Updated July 5, 2026

This is the first instalment of a fixed-point observation of our own visibility (2 days, 40 runs). Our own mentions and citations came out at 0/40 — the pre-action baseline. Of the 13 questions containing "AIO," AI read the term as security measures, a water-cooled cooler, or an autonomic-nervous-system diagnosis, as well as its intended meaning of AI search optimization; the next day the interpretation had shifted (causation not claimed). Research dates: 2026-07-02 / 2026-07-03. Engine: OpenAI Web Search (gpt-4o).

01 / Background

Improving AI-search visibility starts with a baseline: what is the current state? This series observes a fixed set of 20 questions about our own company (Netsujo / Netsujo SIGNAL) at fixed points in time, and publishes the numbers and the observed facts as they are. This first instalment is the pre-action baseline.

As soon as we started measuring, we ran into a problem that comes before counting mentions and citations at all: AI interprets the word "AIO," which we used in our questions, in other ways entirely. If the word for a category does not land correctly, we cannot even get onto the field for that category's questions.

So this first instalment publishes the baseline alongside a check of how "AIO" was interpreted, based on a human review of the full answer text for the 13 (of 20) questions containing "AIO," across 2 days.

02 / Research design

• Population — a fixed 20 questions that do not name us directly (4 groups × 5 questions: company recommendation, problem-solving, purchase decision, target-company questions). This measures our own visibility; it does not evaluate any named competitor.
• Execution — each question run once per day, on OpenAI Web Search (gpt-4o), on 2026-07-02 and 2026-07-03, under identical conditions.
• Recording — question, whether mentioned, cited URLs, candidate competitors, and the start of each answer, saved per group and day. Competitor identification is not asserted automatically.

03 / Main findings

• Across the fixed 20 questions over 2 days (40 runs total), we were mentioned 0 times and our official site was cited 0 times. We publish this as the pre-action baseline and the starting point for ongoing fixed-point observation.
• Of the 13 questions containing "AIO," only 3 were answered with the interpretation "AI search optimization" (close to our intended meaning) as of 7/2. The rest split into security measures (3), an unrelated field such as water-cooled PC coolers or autonomic-nervous-system diagnosis (2), and vague catch-alls like "all-in-one" (5).
• The answer to "what does AIO work typically cost?" turned into an explanation of water-cooled CPU-cooler pricing — a category question answered as if it were about an entirely different market.
• On the re-run the next day (7/3), the number of questions interpreted as "AI search optimization" rose from 3 to 4, and a question that the day before had listed security companies answered in terms closer to our published definition. AI answers vary run to run, so we do not claim causation with our own action.
• The interpretation changed day to day even for the same question. A single measurement cannot establish how a category word is being interpreted, or how we are being treated.

04 / Data

05 / Breakdown

13 questions containing "AIO," classified by a human review of the full answer text, over 2 days.

06 / Analysis

Zero mentions, zero citations is not "failure" — it's a baseline worth publishing. A result of never appearing across 20 questions without naming ourselves looks, we think, like a standard starting state for a company before it has taken action. Our companion study ("Who does AI search actually cite?") found that under a realistic setting citations barely happen at all, and where they do, they spread across 133 domains. We publish this 0/40 as our starting point and re-measure at each milestone under the same conditions; where something changes, we record it as a change, without asserting causation with our own actions.

When a category word is ambiguous, each question gets pulled toward a different market. "Who can I commission for AIO measures?" got a list of major security companies on 7/2; "what does AIO work typically cost?" turned into water-cooled CPU-cooler pricing; "what can an AIO diagnosis tell me?" turned into an explanation of autonomic-nervous-system diagnosis. AIO already has multiple existing meanings — all-in-one PCs and coolers, a security-industry abbreviation, health-related trademarks — and as long as AI defaults to an existing meaning, there is no room for us to appear in answers to category questions that use a new term. This is a structural challenge for anyone trying to define a market with a new term such as AIO, AEO, or GEO.

Interpretation shifted within a day — but we do not claim causation. On the re-run (7/3), the question that had listed security companies the day before answered in terms closer to the definition we expanded public disclosure of that same day (AI search optimization: measures that make a company's information easier for AI search and generative AI to cite and recommend appropriately), and the number of correctly interpreted questions rose from 3 to 4. But AI answers vary run to run even under identical conditions, and the security-context interpretation was still present in 3 questions as of 7/3 — the confusion itself had not gone away. We cannot tell from this observation alone whether the shift was caused by our action, and we check whether the interpretation distribution holds up in the 2026-08 monthly re-measurement.

07 / Limitations of this study

• Each question was run once per day, so the observation is affected by variance.
• The only engine measured was OpenAI Web Search (gpt-4o).
• Interpretation classification is our own manual read, including borderline cases; full raw answers are saved as source data.
• This is a 2-day observation and does not generalise beyond it.

08 / Practical implications

• When defining a business with a new category term (AIO, AEO, GEO, and similar), the starting point is to actually measure how AI interprets that term. Acting on it while it is still misinterpreted just gets buried in a different market.
• For an ambiguous category term, stating the definition clearly on your own site and key pages may need to come before optimising individual pages. Note (2026-07-07): llms.txt is not used by Google Search and does not affect visibility or ranking there (Google, confirmed 2026-06-15) — we no longer recommend putting term definitions in llms.txt for Google Search purposes, and keep it there only for other services or possible future experiments.
• Because AI answers vary day to day, a single measurement is not enough to conclude "we're not listed" or "the interpretation is fixed now" — fixed-point observation under identical conditions is needed to see a trend.
• If you are starting from zero mentions, the basic move — per our companion study's finding that 94% of citations went to companies' own official pages, varying by field — is to build out official pages that answer the question directly.

09 / Reproduction

• Question set — fixed 20 (4 groups × 5), version-managed alongside the scan baseline code.
• Execution — each question run once on OpenAI Web Search (gpt-4o), on 2026-07-02 and 2026-07-03, under identical conditions.
• Recording — question, mention, cited URL, candidate competitor, and answer opening saved per group and day.
• Interpretation classification — for the 13 questions containing "AIO," a human reviewed the full answer text and assigned one of 4 categories (AI search optimization / security / unrelated field / other). Competitor identification is not automated here either.
• Re-measurement — at each action milestone (roughly monthly), run the same questions against the same endpoint and append to the same record.

Related: our companion study, "Who Does AI Search Actually Cite?"