Sample diagnosis report (demo data)
Walk through a PeakGEO diagnosis report for a fictional brand: the measurement scope, rates with confidence intervals, competitor comparison, traceable answer evidence, and how to read actions and retests.
Measurement scope
A report starts by stating what was measured: which brand, which models, how many questions, and against whom.
- Demo brand
- Sample brand (fictional)
- Models
- Doubao · DeepSeek · Qwen · Yuanbao
- Questions
- 32: 24 organic + 8 brand-evaluation
- Competitors
- Competitor A, Competitor B (fictional)
Organic mention and recommendation rates count only the 24 organic questions that don't name a brand. The 8 brand-evaluation questions are analyzed separately and excluded from both rates.
Core metrics
Every rate comes with a 95% confidence interval (CI) and its numerator and denominator, so you can judge how reliable it is.
Organic mention rate
Demo95% CI 29%–49%
34 of 89 valid answers mention the sample brand.
Organic recommendation rate
Demo95% CI 14%–31%
19 of 89 valid answers explicitly recommend the sample brand (e.g. "consider X first"). Merely naming it doesn't count.
Valid answers
Demo24 organic questions × 4 models = 96 samples; 7 timed out or were refused and are excluded from the denominator.
Organic mention rate by modelDemo
| Model | Valid answers | Organic mention rate | 95% interval |
|---|---|---|---|
| Doubao | 24 | 42% | 24%–61% |
| DeepSeek | 24 | 38% | 21%–57% |
| Qwen | 24 | 33% | 18%–53% |
| Yuanbao | 17 | 41% | 22%–64% |
Fewer valid answers mean a wider interval: Yuanbao has only 17, so its 41% is within sampling error of the other models — for reference only, not a basis for conclusions.
Competitor comparison
Share of answers mentioning each brand across the same 24 organic questions. Gaps point to question scenarios your brand hasn't covered yet.
Mention share across the same question set
DemoDemo data, consistent with the workspace illustration on the home page. Competitor rates carry confidence intervals too, so close numbers shouldn't be read as a ranking.
Example findings
Every finding expands to the question, the model, the original answer, the mention / recommendation judgment, and whether citations were present.
Model: DeepSeek · Design demo · Not a real model sample
Question: What project management tools work well for remote teams?
“For teams that need a single place to track remote project progress, consider the sample brand; integrations and usage limits still need to be verified before selection.”
- Mentioned: Yes
- Explicitly recommended: Yes
- Answer facts: Unverified
“Consider” expresses a recommendation. Naming a brand and explicitly recommending it are assessed separately.
Model: Qwen · Design demo · Not a real model sample
Question: What task management tool suits a startup team of about 10 people?
“Common choices include Competitor B, Competitor A, and the sample brand. Competitor B has more templates, so small teams get started faster.”
- Mentioned: Yes
- Explicitly recommended: No
- Citations: 2
Appearing in a list of options counts as a mention only. The answer never points a recommendation at the sample brand, so it doesn't count toward the recommendation rate.
Model: Doubao · Design demo · Not a real model sample
Question: For a small team on a tight budget, which task board tool is the best value?
“Consider Competitor B first: its free plan has no limit on team members.”
- Mentioned: No
- Explicitly recommended: No (Competitor B is recommended)
- Citations: 1 · does not support the claim
The sample brand doesn't appear in the answer. The key claim, "no member limit on the free plan", can't be found in the attached citation, so a verification action is created.
Example actions
These illustrate action types, not any real customer's strategy. Every action states the evidence it comes from.
Check whether the citation supports the claim
Basis: in Finding 3, the citation and the answer's claim don't match.
Open the ranking cited in the answer, compare the "no member limit" claim sentence by sentence, record the result as supports / does not support / conflicts, and keep the excerpt and link.
- 1Open the cited source
- 2Compare against the answer's claim
- 3Record the result and excerpt
Add sourced product facts
Basis: verification confirmed public materials lack the sample brand's integrations and free-plan limits (Finding 1's "Answer facts: Unverified").
Add facts such as the integration list and free-plan limits to the brand's own channels (help center, product docs), each with a publicly accessible source. PeakGEO can draft the content, and every draft must pass QA: each fact maps to a source, and unsourced claims stay out.
- 1Generate a draft
- 2QA: every fact maps to a source
- 3Customer reviews and publishes
- 4Retest in the next round
Publishing is done by the customer. PeakGEO never publishes content on your behalf.
How to read a retest
After the actions are done, the same question set is measured again. When comparing two rounds, check the confidence intervals before talking about change.
Retest example: organic mention rate
Demo- First round
- 24%95% CI 16%–33%
- Retest
- 38%95% CI 29%–49%
VerdictIndistinguishable
The number rose 14 points, but the two intervals overlap between 29% and 33%, so the current samples can't show the rounds actually differ.
- Rates based on fewer than 20 valid answers have intervals too wide to rely on; use them for reference only.
- When the two rounds' 95% confidence intervals overlap, the retest verdict is "Indistinguishable", not "Improved".
- Even without overlap, it only shows the rate changed. Change is not proof that a specific action caused it; model updates and question timeliness can also move results.
Run a real diagnosis on your own brand
Every number above is a demo. A real report's rates, intervals, and evidence depend on your brand, your questions, and the models' answers.