Measurement notes · October 2026

1. What changed in the measurement in October

Perplexity: answers open with a search log

Since the scan of 2 October, every Perplexity answer opens with the engine's own search log (its query and the sources it found). Scoring reads the opening of an answer, so in the scheduled scans of 2 and 9 October the prominence, sentiment and presence of Perplexity were scored on that log, not on the answer. This was corrected on 9 October. Those two scans are marked * in reports and are not used to call a rise or a fall; the first scheduled scan after the correction is the new baseline for Perplexity.

Google AI Overview: answers that were not retrieved used to count as "absent"

For many questions Google generates the AI Overview after the results page has loaded, and its text has to be fetched with a second request. Before 11 October the two requests shared one 30-second limit; when it ran out, or the text could not be fetched, the answer was stored as empty and scored as the brand being absent. Since 11 October each request has its own limit, a failed one is retried once, and every answer is recorded as one of three outcomes: answered, no AI Overview shown, or not retrieved. Because more answers now come back, the Google AI Overview reading from the first scan after 11 October is not fully comparable with earlier scans; reports say so once.

Perplexity lists fewer sources

Also since early October, Perplexity lists far fewer sources per answer: about 5, against 25–30 before (measured on one Vietnamese question panel). This is the engine's own change. Perplexity's citation counts and source lists before and after that date cannot be compared with each other.

2. When a number counts as having moved

A report calls a change a rise or a fall only when these rules are met.

1. Only scheduled scans are compared: the same questions, the same engines, the same time of the week. Scans run by hand are listed separately and never compared.
2. A change of the overall score within ±3 is sampling noise. The report prints the number and does not call it a rise or a fall.
3. A marked scan — scored on the wrong text, an engine that did not return most of its answers, or a provider that changed its model — is not used for conclusions. The next valid scan is the new baseline.
4. A question that "has" or "does not have" the brand in one scan is not yet a change. Asked again a few hours later, the same question can come out differently: in one check, 11 questions were asked three times in one evening and 3 of them did not give the same result each time. Only a result that repeats across scans is treated as stable.
5. Nobody adjusts a measured number by hand. When the instrument is at fault we repair the instrument, keep the recorded number and mark it.

3. How a number is produced

Every scan passes through four layers. Only one of them contains the judgement of an AI model; scoring is done by rule.

Asking

A fixed panel of questions in Vietnamese is put to five engines: ChatGPT, Claude, Gemini, Perplexity and Google AI Overview. The first four are asked through the model's programming interface, so an answer can differ from the app on a user's phone. Google AI Overview is taken from the real search results page.

Retrieving

Answers are stored in full. An engine's own working log, where there is one, is removed before scoring.

Reading

A language model reads the opening of each answer and records four things: whether the brand is mentioned, how prominently, in what tone, and which other brands are mentioned. This is the only layer with a model judgement; it is checked against a set of answers labelled by people.

Scoring

AIGVR is a fixed formula: presence 30%, prominence 25%, share of voice against competitors 20%, sentiment 15%, citation 10%. No model takes part in this layer, so the score can be recomputed from the same data.

In the weekly report, Part 1 is the measurement. Part 2 (interpretation) and Part 3 (suggestions) are written by AI; every claim is checked against the scan and a sentence that cannot be verified is removed. Neither part changes the numbers in Part 1.

On "decision models": the reading layer currently uses a general-purpose conversational model. A newer kind of model, which returns probabilities instead of text, is appearing; it is cheaper and measures "not sure" more finely. We are evaluating it. Replacing the model in the reading layer is a change of method: it happens only when the new model reaches the required agreement with the answers labelled by people, and it is declared in the report for that period.

4. A new method: paired-predicate verification

Not yet applied to Vietnamese-language projects

Presence counts the brand's name; it does not measure what AI "believes" about the brand. A brand can be mentioned often and still be described wrongly.

This method takes a claim from the brand's own published material, writes its direct negation, and puts each sentence to the engine in two separate requests: is this statement correct? The engine may answer only yes, no or unsure, with one sentence of basis. The pair of answers gives one of four outcomes. Each set is asked at least twice; when the two rounds differ, the fifth outcome is recorded.

• The two sentences are asked separately. In one conversation the engine keeps its two answers consistent on purpose.
• Every round includes control sentences unrelated to the brand, half of them known to be false. An engine that agrees with a false one has all its results for that round discarded.
• A claim that carries two facts is split, and each fact is asked on its own.

The method gives no "accuracy score". An accuracy score needs a referee who knows the truth, and the truth about a company is in the material the company itself publishes, not with us and not with a model. The method observes two things only: whether AI says what the brand publishes, and whether AI is consistent with itself.

Limits: it does not decide what is true; it measures answers obtained through the model's programming interface; and a brand has a dozen or so claims, enough to show which sentences need a verifiable source, not enough for a general conclusion about an engine.

The method was developed and published by NeuronSpark (观澜智库, Hong Kong); MemeCMO uses it under licence. It has been run on a demonstration brand in Hong Kong, in Chinese. Before it is applied to a Vietnamese-language project it needs Vietnamese questions and controls, and a list of claims taken only from the brand's published material and confirmed by the brand.

5. Two things that let the measurement reflect your team's work

MemeCMO does not publish content to any AI. AI reads public web pages only. What is logged or edited on the platform (the brand profile, drafts, the record of actions) changes our records and suggestions and nothing else; for AI to see something, it has to be published on your website or in the media.

• Log every piece of content you publish (link and date). Without that record a report cannot connect an article to a change in what AI answers.
• Confirm the brand's reference facts. The same list is the claim set for the paired-predicate method.