Skip to content

Methodology

How ReleVantage improves AI search visibility

Six stages, run in order, across ChatGPT, Google AI, Gemini, Perplexity, Claude, and Copilot. Measurement comes first so that every later claim can be checked against a recorded baseline.

01

Baseline measurement

Before anything is changed, we record how each platform describes you today. The same prompts are run more than once per platform, full answer text is stored, and every mention, citation, and factual error is logged. That record becomes the reference point for every later claim of progress.

  • Answer text captured verbatim per platform and per run
  • Mentions, citations, and cited domains logged separately
  • Factual errors recorded with the likely source of each
  • A written scoring rubric so later runs are comparable
02

Prompt and competitor mapping

We build the prompt set from real buyer language rather than keyword volume, then map who currently occupies each answer. Prompts are grouped so that category, comparison, factual, and constraint driven questions can be reported apart from one another.

  • Category, comparison, factual, constraint, and branded prompt groups
  • Named competitors tracked per prompt group
  • The sources each competitor relies on identified
  • A frozen core set, with new prompts tracked as an additional group
03

Technical and entity correction

Assistants cannot use what they cannot crawl or verify. This stage repairs the foundation: indexation, structured data, consistent naming, and agreement between your site and the profiles and directories that describe you.

  • Crawlability, indexation, duplication, and rendering issues resolved
  • Structured data aligned to what is visible on the page
  • One canonical description, category set, and service list
  • Hours, addresses, coverage, and contact details made consistent everywhere
04

Evidence asset creation

We write or restructure the pages that answer the mapped prompts, in a form a model can quote without risking a wrong claim. Answer first structure, self contained passages, explicit qualifiers, and plain statements of what is and is not true.

  • Existing high-intent pages restructured for extraction
  • New pages only where a real buyer question has no home
  • Constraints, limits, and exclusions stated rather than implied
  • Facts sourced and dated so they can be maintained
05

Legitimate third-party corroboration

Models look for agreement between independent sources. We identify the publications, registries, marketplaces, and directories that are actually cited in your category, then pursue accurate coverage on them through legitimate outreach and correction of existing records.

  • Cited source categories identified from your own baseline data
  • Existing records corrected before new placements are pursued
  • Outreach for roundups, comparisons, and expert commentary
  • Source diversity tracked so presence does not rest on one domain
06

Repeated testing and reporting

The prompt set is re run on a fixed schedule and scored the same way each time. Reporting shows the baseline, the current run, what changed, what work preceded it, and what is scheduled next. Where a movement cannot be explained, we say so.

  • Fixed schedule, fixed wording, multiple runs per prompt
  • Change reported against baseline rather than as a single score
  • Unexplained movement flagged rather than claimed as a result
  • Next actions prioritised in a working session, not a static PDF

What we measure, kept separate

A single visibility score hides the useful detail. These eight measures are recorded and reported independently.

Visibility

The share of tracked prompts where the brand appears in the answer at all, reported per platform.

Mention position

Where in the answer the brand appears. Named first is a different outcome to a closing list.

Sentiment

How the model characterises the brand, including framing such as premium, limited, or expensive.

Accuracy

Whether the facts stated are correct. Each error is logged individually and traced to a source.

Citations

Which URLs are attributed, split between your own pages and third-party pages that discuss you.

Source diversity

How many independent domains support your presence, because reliance on one source is fragile.

Qualified traffic

Sessions from assistant referrers judged on engagement and conversion rather than raw volume.

Business outcomes

Enquiries, bookings, and demos associated with those sessions, reported with attribution limits stated.

Results vary and placements are not guaranteed. Read more in how to measure AI search visibility, or see how the methodology is applied by industry.

Methodology questions

Why start with measurement instead of content?

Without a baseline there is no way to tell whether a later change came from the work or from a platform update. The baseline also shows which problems are factual errors, which are entity confusion, and which are genuine absence, and those need different fixes.

Why are visibility and accuracy scored separately?

A brand can be mentioned frequently and described incorrectly. Combining the two into one score hides the more damaging problem, so each answer is scored on both.

How often is the prompt set re run?

Weekly by default, with a shorter interval for a smaller priority group in volatile categories. Each prompt is run more than once because generated answers vary between runs.

Can improvement be attributed directly to the work?

Only in part. We report what changed and what work preceded it, and we state where a movement cannot be explained. Platforms update independently of anything we ship.

Does this methodology guarantee a placement?

No. Results vary with crawlability, existing authority, competition, source coverage, and each platform's refresh cycle. ReleVantage does not guarantee a specific placement or timeline.

Get in touch

Have a GEO or AISEO question? Send us a message and we will get back to you within one business day.