Short answer
- Never accept a single cherry-picked screenshot or an unverified percentage claim ('visibility up 83%') without the prompt list and multi-run logs.
- Separate four distinct metrics: (1) Brand Mention, (2) Positive Top-3 Recommendation, (3) Grounding Source Citation, and (4) Attributed Inquiry/Conversion.
- Filter out negative or exclusion mentions — being named as 'not suitable for consumer skincare' is a defect to fix, not a visibility win.
Why single screenshots and aggregate percentages mislead
When evaluating GEO agencies for a beauty brand or cosmetics manufacturer, you will frequently see claims such as "we increased a beauty brand's AI recommendation rate from 7% to 83%" or a single cropped screenshot of ChatGPT. Because LLM outputs vary across runs and prompt phrasing, an agency can test 50 variations, screenshot the one lucky run, and call it a case study.
The 4-Layer BG-STD-001 Measurement & Attribution Standard
| Metric Layer | Exact Definition | Common Agency Reporting Trap | How BeautyGEO Verifies It |
|---|---|---|---|
| Layer 1 · Brand Mention | Your brand, product or factory name appears anywhere in the AI answer text | Counting neutral lists or negative/exclusion statements ('X is only for B2B factories, not consumer brands') as positive visibility | Every capture is tagged as primary_recommendation, secondary_mention or negative_or_exclusion; exclusions are treated as defects to remediate |
| Layer 2 · Positive Recommendation & Slot Share | The AI actively recommends your product or supplier in Slots #1–#3 for an unbranded buyer/shopper prompt | Changing prompts month to month or running only 1 sample per prompt | Fixed cohort of 30–50 prompts re-tested across 3 independent runs per engine every month |
| Layer 3 · Grounding Source Citation | The AI attaches a clickable footnote link to your canonical site or a verified third-party review/directory | Taking credit when the AI links to a generic marketplace page that lists 20 competitors | Logs exact cited URLs, separating owned HTML pages from retailer feeds and earned media |
| Layer 4 · Attributed Visits & Inquiries | Sessions arriving from AI referrers plus self-reported 'How did you find us?' leads and orders | Reporting 'millions of AI impressions' with zero website traffic or RFQs | First-touch referrer tracking (chatgpt.com, perplexity.ai, gemini.google.com) + form attribution + CRM verification |
Five things to verify before signing any Beauty GEO agency
- Ask for the verbatim prompt cohort and 3-repeat transcripts — not just a summary chart.
- Check whether they separate affiliated pilots from non-affiliated open benchmarks (such as our open CC BY 4.0 79-brand Beauty AI Shelf Index).
- Test their cosmetic regulatory fluency on US FDA MoCRA, EU 655/2013 and China NMPA efficacy boundaries.
- Inspect their own website's technical hygiene (
/llms.txt,Organization,ProductandDefinedTermSetJSON-LD). - Require published pricing and clear exit terms where you keep every page and schema asset built.