How we measure visibility and demand in AI assistants

Starting point: an observation, not demonstrated improvement

In a manual test on 1 November 2026, ChatGPT using GPT-5.6 Sol, medium reasoning, a temporary conversation and no personalisation cited our website when asked about Multiconversion and correctly reproduced our commercial contact details. A separate query asking for five Valencia agencies for SEO/GEO and Google Ads did not include Multiconversion.

The previous generic test used a different model. These observations are therefore not a controlled before-and-after comparison or a representative market measure. They also do not establish that the assistant has incorporated the latest website changes.

A repeatable protocol

The panel contains 16 Spanish-language questions for Spain: four about hotels, four about ecommerce, four about education and four about the brand. Of the 12 unbranded questions, nine ask for suppliers and three are informational (H04, E02 and D04). Wording and identifiers are preserved. Intent classification was added after observing the initial panel, in version 2026-11-01.2; it does not turn this tracking into a causal before-and-after comparison.

The design calls for two independent runs per question in each assistant at the start and on days 30, 60 and 90. This means 128 expected responses per measurement when coverage is complete. Pending or failed runs are not reported as negative answers.

We record date, application or API, visible model and version, language, market, web-search availability, personalisation and the complete answer with sources. Sessions must be independent. A change of model or channel is flagged and analysed as a separate series.

What we are testing

The work combines sector-specific commercial content, cases with explicit scope, consistent company information, buying guides and a diagnostic connected to the sales process. Accessible, stable HTML for people and agents is also maintained.

We will log publication dates and affected pages. Subsequent changes may coincide with search-engine, model or third-party website updates: this observational tracking cannot by itself attribute all improvement to one intervention.

From an answer to a commercial opportunity

We combine an identifiable visit source with the contact’s declared source and sales progress. We deduplicate before counting opportunities. A recommendation without a click can help a buyer even if it leaves no identifiable referrer.

Cost per qualified opportunity uses allocated programme costs and AI-origin opportunities defined in advance. Assisted opportunities are reported separately. We do not publish contacts’ personal data or confidential client information.

Four levels of measurement

Access

HTML availability, crawling, indexing and content retrieval. A technical prerequisite.

Recommendation

Valid answers to the nine supplier-selection questions proposing Multiconversion ÷ valid answers to those same questions, using the same assistant and settings. Informational and branded questions are excluded.

Citation

Valid answers linking to multiconversion.com ÷ valid answers in the group analysed. We check whether the source supports the claim.

Business

Qualified enquiries, proposals, contracts and margin. Detected origin, declared origin and assisted contribution are separated.

We will publish numerators and denominators, coverage by assistant and errors. A mention, a citation and a recommendation are distinct categories. The three informational questions are assessed for mentions, citations and accuracy, separately from commercial recommendations; a useful answer may name no suppliers. The four branded questions also form a separate group.

What this protocol does not guarantee

Allowing crawling and maintaining a correct llms.txt does not guarantee indexing, citations or recommendations. Google states that Search ignores llms.txt files and relies on its search foundations. Providers distinguish between crawler purposes; we review their official documentation.

The four assistants are distinct surfaces. An API or tracking-tool answer is not presented as an identical test of the consumer application. The panel measures a defined question sample, not total demand share.

• Google Search
• OpenAI
• Anthropic
• Perplexity

The 16 panel questions

Questions will be run in Spanish to preserve the same sample. The downloadable file contains the questions and planned settings, without internal account data.

Download the question panel (JSON) ↓