Intro
Most brands ask “are we visible in AI?” before they have a way to answer it. A baseline is the way.
A baseline is the first three weeks of every Capston Core engagement. It is not a report, not a dashboard, not a one-off audit. It is a locked measurement frame — prompt set, competitor set, engine set, capture method — that every later retest will be compared against.
This page walks through what running a baseline actually looks like, step by step, with the six concrete outputs it produces.
What a baseline produces
A Capston Core baseline produces six locked outputs:
- A prompt set of 40 to 80 prompts, agreed with the brand and locked.
- A defined competitor set used for every comparison.
- AI answer captures stored with model version and date metadata.
- A citation map showing which URLs each engine reuses, per prompt.
- Fact-accuracy notes flagging correct, wrong, or missing claims about the brand.
- A first scorecard snapshot across the eight visibility dimensions.
Once these six exist, the brand has a comparable position. Without them, every later number is an impression.
Step 1: Co-design the prompt set
The prompt set is the foundation. If it is wrong, everything downstream is wrong.
Capston works with the brand to assemble 40 to 80 prompts split across four intent buckets:
- Discovery — broad queries that surface a shortlist (“best wellness retreats in the Alps”)
- Comparison — head-to-head and “best of” queries
- Trust — reputation, review and accuracy queries
- Conversion — branded queries where the answer drives or breaks a booking
The brand contributes the prompts that match how their real customers ask. Capston contributes the prompts that pressure-test commercial risk. Both lists are merged, deduplicated, and locked.
Locked matters. Adding prompts mid-cycle invalidates the comparison. Anything new is queued for the next retest.
See the Capston Core methodology for how the prompt set feeds the rest of the five-stage process.
Step 2: Lock the competitor set
The second locked artefact is the competitor set.
Capston works with the brand to name the five to twelve competitors that the score will track. Two rules:
- Include the competitors the brand worries about.
- Include the competitors AI engines already name, even when the brand does not see them as direct rivals.
The second rule is the one brands miss. AI engines often surface adjacent operators, aggregators or intermediaries that the brand has not benchmarked against. Those entities show up in answers and steal share regardless.
The competitor set is locked alongside the prompt set. Both are versioned. A new competitor entering the market is queued for the next retest.
Step 3: Capture AI answers
With the prompt set and competitor set locked, capture begins.
Every prompt is run across the relevant AI engines. The default set is ChatGPT, Perplexity, Google AI Overviews and Gemini, adjusted by market. Each answer is captured with:
- The exact prompt
- The engine and model version
- The date and time
- The full answer text
- Any cited URLs or sources surfaced by the engine
These captures are stored in the evidence layer. Nothing is paraphrased, summarised or interpreted at this stage. The capture is raw. Interpretation happens later, against the raw record, so any claim can be traced back.
This is the principle behind the AI answer evidence layer: every score, every recommendation, every retest delta points back to a dated capture.
Step 4: Build the citation map
From the captures, Capston builds a citation map.
The citation map answers four questions per prompt:
- Which URLs did the engine cite?
- Does the brand’s own domain appear in the cited sources?
- Which third-party sources dominate (review sites, OTAs, press, aggregators)?
- Which competitor domains are reused across prompts?
The output is a per-prompt table and a portfolio-level rollup. Brands almost always discover that their own domain is cited far less than they expected, and that a small set of third-party sources concentrate most of the citation share.
The citation map is what turns “we are not visible” into a specific, addressable problem.
Step 5: Note fact accuracy
The fourth diagnostic layer is fact accuracy.
Capston reads each answer and flags every factual claim about the brand:
- Correct — the engine described the brand accurately.
- Wrong — the engine stated something the brand can disprove (wrong location, wrong amenities, wrong price band, outdated ownership, retired offer).
- Missing — a claim that should be present is absent (signature service, key differentiator, accreditation).
Each note is linked to the capture it came from. No claim is corrected from memory. Wrong facts are the highest-priority fix in stage three of Capston Core, because they actively misinform customers in the answer surface itself.
Step 6: Produce the scorecard
The final baseline output is the first scorecard snapshot.
The scorecard rates the brand across eight dimensions: brand presence, answer position, citation share, source quality, competitor dominance, fact accuracy, sentiment, and commercial risk. The methodology is detailed in AI visibility scoring.
The first scorecard is not a verdict. It is a position. It tells the brand:
- Where it sits today, per dimension, per engine.
- Which competitors lead, on which prompts.
- Where the biggest gap concentrates.
- Where the easiest gain lives.
- Where commercial risk is most exposed.
For hospitality brands, the snapshot maps onto the Capston Hospitality Scorecard, with its vertical-specific risk surfaces around OTA capture and intermediary substitution.
The scorecard closes the baseline. Stages two through five of Capston Core — diagnose, fix, verify, monitor — all read from it.
Timeline and who does what
A baseline runs across three weeks.
- Week 1. Prompt set drafted by Capston, reviewed and amended by the brand, locked. Competitor set agreed and locked. Engine set confirmed for the brand’s markets.
- Week 2. Captures run by Capston across the locked engine set. Citation map built. Fact-accuracy notes drafted by Capston, reviewed by the brand for claim verification.
- Week 3. Scorecard assembled, calibrated, and reviewed against the Capston QA standards. Baseline readout delivered, with the biggest gap, easiest gain and biggest risk surfaced explicitly.
The brand contributes roughly four hours total across the three weeks: one hour on prompts, one on competitors, two on fact review. Everything else is run by Capston.
How this fits into Capston Core
The baseline is stage one of five.
Stage two — diagnose — reads from the citation map and the fact-accuracy notes. Stage three — fix — addresses the highest-priority gaps surfaced in the scorecard. Stage four — verify — reruns the locked prompt set against the same engines to measure delta. Stage five — monitor — establishes the retest cadence, typically quarterly.
Without a properly locked baseline, stages two through five lose their reference frame. With one, every later action has a number attached.
→ Back to Capston Core
FAQ
How long does a baseline take?
Three weeks end-to-end. Week one locks the prompt and competitor set, week two runs captures and builds the citation map, week three produces the scorecard.
How much time does the brand need to commit?
Roughly four hours across three weeks: prompt review, competitor review and fact-accuracy review. Capston handles capture, mapping and scoring.
Why lock the prompt set instead of adding prompts over time?
Because the baseline only has value as a comparison frame. Adding prompts mid-cycle changes the denominator and invalidates the delta. New prompts queue for the next retest.
Can a baseline be run without naming competitors?
No. Without a locked competitor set, citation share, competitor dominance and answer position cannot be measured consistently. The competitor set is non-optional.
Final CTA block
Run a Capston AI visibility baseline.
Start a baseline
Read the methodology