In this blog
The first time a brand team sees systematic AI-visibility data, the reaction is always the same: disbelief, then reorganization of the roadmap. Not because the news is uniformly bad — but because it is specific in ways brand tracking never was. This is the methodology behind that specificity, as we run it with Adobe Brand Visibility — the instrument panel; this post is the flight manual.
Step 1: Build the prompt portfolio — the measurement's DNA
Everything downstream inherits the portfolio's quality. Build it from buyer language, structured into clusters:
- Category prompts — "best [category] for [segment]" — the consideration-set formers
- Comparison prompts — "[you] vs [competitor]" and "[competitor] vs [competitor]" — closest to purchase, highest stakes
- Validation prompts — "is [brand] good for [use case]", "[brand] problems" — where sentiment lives
- Problem prompts — the unbranded questions your solution answers — where challengers steal futures
Source from sales-call transcripts, support themes, community threads and search queries rephrased as questions. Then weight by commercial value: the prompt behind your highest-margin conversation deserves more attention than a long-tail curiosity. A portfolio of 75–200 weighted prompts covers most brands with statistical stability; the discipline mirrors the prompt-set strategy in LLMO work, because it is the same DNA serving measurement instead of remediation.
Step 2: Score four metrics, not one
| Metric | What it captures | Why counts alone mislead |
|---|---|---|
| Share of voice | Presence rate across the portfolio vs competitors | The baseline — but presence ≠ endorsement |
| Recommendation rate | How often you are recommended, not just named | "Mentioned as the dated option" inflates SoV |
| Sentiment & context | The qualities and caveats attached to your name | Context drives strategy more than frequency |
| Citation sources | Which sources feed answers about you | Reveals the levers — whose content to influence |
The fourth metric is the secret weapon: discovering that a review site and one competitor comparison page feed most category answers converts a vague visibility ambition into a named influence target list.
Step 3: Segment ruthlessly — averages lie
A single global score hides everything actionable. Cut the data three ways:
- By engine — sources and synthesis differ across ChatGPT, Perplexity, Gemini and AI Overviews; strong on one, absent on another is the normal finding, and your buyers do not distribute evenly across them
- By market and language — multi-market brands routinely find a healthy US narration beside a damaged one elsewhere; the product's segmentation exists because this is where multi-nationals get surprised
- By topic cluster — owning awareness prompts while losing comparison prompts is a specific, fixable pattern (strong brand, weak proof content), invisible in the average
The reporting rule: every rollup must be decomposable in two clicks, or someone will make a decision on an average that no segment actually exhibits.
Step 4: Cadence and alerting
Scan cadence follows volatility — fast-moving categories (AI tooling, fintech) warrant frequent measurement; stable ones can breathe slower. The architecture that works: scheduled scans for trends, alerts for inflections — a competitor's sudden surge in comparison prompts, a sentiment drop after coverage, a money-conversation falloff. Alerts convert the program from quarterly wallpaper into an early-warning system; the inflections, not the levels, are where response windows live.
Step 5: Report position, not plumbing
Leadership does not act on mention counts; it acts on competitive movement:
- The displacement chart — your share vs. named competitors by cluster, quarter over quarter; the single most-requested view
- The context digest — the three qualities AI most attaches to your brand, and whether they are the three you want
- The wins/losses ledger — conversations gained, conversations lost, and to whom
- The influence targets — sources feeding negative or competitor-favoring answers, ranked by leverage
Connect movement to action explicitly: visibility gains following content operations prove the loop, as programs like HDFC's AI-search work demonstrate — 940+ answer keywords is a measurement claim backed by exactly this instrumentation.
The methodology failure modes
- Demo-prompt syndrome — three cherry-picked prompts on a call proving whatever the presenter wanted
- Unweighted portfolios — the long tail drowning the money conversations in the average
- Count worship — celebrating mentions while sentiment quietly curdles
- One-engine myopia — optimizing the surface you tested, not the surfaces buyers use
- Measurement without an operator — dashboards inspiring nobody's next action; pair the instrument with the content loop or it decays into wallpaper
Getting started
Two weeks to a working baseline: portfolio construction (week one), first full scan and segmentation (week two), findings deck with displacement chart and influence targets at the end. From there the cadence runs itself, and the findings fund the operation. The traditional-monitoring comparison covers how this slots beside your existing brand trackers; pricing covers the instrumentation cost. For the baseline run by the team that operates this loop end to end, talk to a consultant — the first scan's surprises are free of charge.
Frequently asked questions
How do you measure brand visibility in AI search?
Define a weighted portfolio of category prompts, query the major engines systematically, and score results on brand presence, recommendation strength, sentiment and cited sources — trended over time and benchmarked against competitors.
How many prompts should a visibility program track?
Enough to cover your money conversations with statistical stability — typically 75–200 prompts across category, comparison, validation and problem clusters, weighted by commercial value rather than treated equally.
How often should AI visibility be measured?
Scan cadence follows category volatility: fast-moving categories warrant frequent scans, stable ones less. The pattern that works: regular scheduled scans for trends plus alerting for significant shifts between them.
What is a good share of voice in AI answers?
There is no universal benchmark — position relative to your competitive set is the meaningful read. The actionable questions: are you present in the conversations you monetize, trending up, and displacing competitors in comparison prompts?
Do different AI engines give different brand answers?
Substantially — engines differ in sources, recency and synthesis behavior, so brand narration varies across them. Engine-level segmentation is essential; a strong position on one surface can mask absence on another your buyers use.