Method in brief
- Prompts: 21
- Prompt types: category, comparison, location, bottom of funnel and persona
- Engines: ChatGPT, Claude, Gemini, Perplexity, Grok, Mistral and DeepSeek
- Mode: augmented
- Web search: on
- Runs per prompt: 3
- Responses per measurement: 441
21 prompts × 7 engines × 3 runs = 441 responses.
Context
The company had technical content about installation, equipment and energy output. Even so, the brand appeared irregularly in questions about comparing suppliers, choosing a solution and requesting a quote.
The information existed, but was scattered across commercial pages, technical articles and loosely connected documentation.
Baseline measurement, T0
- Date
- 20 January 2026
- Citation rate
- 9.1%
- Mention rate
- 25.6%
- Recommendation rate
- 13.8%
- Share of voice
- 6.4%
- Engines recommending
- 2 of 7
- Average position
- 5.2
Reading the baseline
The brand was recognised in generic questions, but lost presence in comparison, location and decision prompts.
The engines found technical information, but struggled to turn it into a clear, supported recommendation.
Work carried out
Structure
January 2026
- Reorganised solution pages.
- Separated installation types clearly.
- Reviewed pages by customer profile.
- Improved structured data.
- Reorganised information by region.
Information
February 2026
- Created pages about comparing quotes.
- Explained the installation process.
- Published information about maintenance.
- Set out selection criteria.
- Added frequently asked questions.
- Published content on eligibility and limitations.
- Improved the factual, quotable elements.
Notability
February and March 2026
- Reviewed external sources.
- Aligned the company description.
- Strengthened sector and local references.
- Corrected divergent information.
Authority
February 2026
- Clarified technical responsibility.
- Named authors.
- Added sources.
- Reviewed technical claims.
- Provided proof of experience in residential projects.
Links
January to March 2026
- Linked solution, region and process.
- Linked technical articles to decision pages.
- Built paths towards requesting a quote.
Final measurement, T1
- Date
- 5 May 2026
- Citation rate
- 31.7%
- Mention rate
- 52.2%
- Recommendation rate
- 38.3%
- Share of voice
- 18.5%
- Engines recommending
- 6 of 7
- Average position
- 2.6
Observed change
- Date
- 20 January 2026
- Citation rate
- 9.1%
- Mention rate
- 25.6%
- Recommendation rate
- 13.8%
- Share of voice
- 6.4%
- Engines recommending
- 2 of 7
- Average position
- 5.2
- Date
- 5 May 2026
- Citation rate
- 31.7%
- Mention rate
- 52.2%
- Recommendation rate
- 38.3%
- Share of voice
- 18.5%
- Engines recommending
- 6 of 7
- Average position
- 2.6
| Metric | T0 | T1 | Change |
|---|---|---|---|
| Citation rate | 9.1% | 31.7% | 22.6 pp |
| Share of voice | 6.4% | 18.5% | 12.1 pp |
| Recommendation rate | 13.8% | 38.3% | 24.5 pp |
| Engines recommending | 2 of 7 | 6 of 7 | up 4 |
| Average position | 5.2 | 2.6 | closer to the start |
Result
Over the period measured:
- Citation rate went from 9.1% to 31.7%.
- The change was 22.6 percentage points.
- Share of voice went from 6.4% to 18.5%.
- The brand went from recommended by 2 engines to 6.
- Average position went from 5.2 to 2.6, meaning the brand began to appear closer to the start of the answers.
Conclusion
Having technical information does not mean that information is organised in a quotable way.
The change tracked in this case followed from making the relationships between solution, process, region, selection criteria and technical responsibility explicit.
Methodology and definitions
T0 and T1 use exactly the same prompts, engines, mode, web-search setting and number of runs, and the same definitions and formulas. Without that correspondence the change would not be computed.
- Unit of observation. Case × measurement × prompt × engine × run.
- Citation rate. Responses with a qualifying citation ÷ total valid responses × 100.
- Mention rate. Responses with a mention ÷ total valid responses × 100.
- Recommendation rate. Responses with a recommendation ÷ total valid responses × 100.
- Share of voice. Recommendations of the brand ÷ total recommendations of every brand observed × 100.
- Engine recommending. An engine with a recommendation rate of 25% or higher across the prompts in the case.
- Average position. Simple mean of the brand's ordinal position, counted only in responses where it appears. A lower figure is better.
Limitations
- Results apply only to the prompts, engines, mode and period measured. They do not extrapolate to other questions or dates.
- Results in AI engines vary by version, mode, location and moment of collection.
- A citation is not an endorsement, and a recommendation does not measure service quality.
- No effect on sales, revenue, enquiries or pipeline is claimed, because none was measured.
Nature of the evidence
Baseline and final exports from the Visibility Tracker, with a record of the responses, retrieved sources and the brand's position.
Back to all cases · Cross-engine consistency of AI visibility for Portuguese consumer-service brands, 2026