REAL MODEL RESPONSES · 22 OCTOBER 2026
One sentence changed.
A better answer.
AI said our launch fee covered five brands. We found the ambiguity, changed one sentence, and ran the question again. The same model then stated the price per brand.
ACTUAL GPT-4.1 NANO ANSWERS
Five brands?
One brand.
BEFORE · WORLD A · REPEAT 1
“The initial setup cost is €1,499, which covers the first five brands, including up to 100 starter articles.”
The model turned a limited offer into a five-brand package. Our package covers one brand.
AFTER · WORLD A · REPEAT 1
“The initial setup fee is €1,499 per brand, covering up to 100 starter articles and the first month of Growth.”
The price is now stated per brand. Both revised GPT answers made this distinction.
Exact excerpts from recorded API answers. Same model, buyer question, instructions, and general article. We changed only the offer wording. Two repeats per version; this is a small exploratory check.
THE GEO PERSONA
A founder.
With real constraints.
WHAT THEY NEED
A small SaaS founder needs 30 articles each month. Their recurring budget is €500 before VAT. They can check facts and approve content. They cannot manage an editorial team.
WHAT AI MUST EXPLAIN
The recurring price. The separate launch fee. Who approves the work. How publishing works. Whether images cost extra. We kept this buyer question fixed throughout the experiment.
THE MODEL WORLDS
One offer.
Three possible moves.
The existing offer facts, supplied directly to the model.
The offer plus a hypothetical article about AI content trends. This is where we selected the before-and-after pricing check.
The offer plus a hypothetical guide to the buying details. Both GPT and Mistral named the guide in both answers. GPT still made a five-brand pricing error in one answer.
The offer plus a hypothetical comparison. Both GPT and Mistral named it in both answers. It was supplied in the prompt, not published or found online.
GPT and Mistral each completed eight answers: two per world. GPT recommended the offer in all eight. Mistral recommended it in six. Adding a source did not establish a winning visibility intervention.
THE SELECTED INTERVENTION
Fix the ambiguity.
Keep the offer.
We chose a small edit to the offer itself. It directly addresses the observed error. It needs no new article or external placement.
ORIGINAL SOURCE TEXT
“Founding Launch is €1,499 once for the first five brands;”
EXACT REPLACEMENT TESTED
“Founding Launch costs €1,499 per brand for each of the first five customers; the launch fee applies to every founding customer and is not waived;”
All remaining source text stayed the same. The replacement was tested in World A. We selected it for its direct fit to the error and small scope; financial ROI was not measured.
WHAT CHANGED IN THE ANSWERS
Check the fix.
Keep the exceptions.
Before: one of two World A answers said the fee covered five brands. After: neither did. Both revised answers explicitly said €1,499 per brand. Recommendations stayed positive in both versions.
Before: both World A answers misread the founding terms. One implied a fee waiver; the other implied five-brand coverage. After: neither repeated those errors. Budget framing remained confusing in the revised answers.
Before: one of two World A answers rejected the offer by mixing setup cost with the recurring budget. After: both stated that €499/month fits. This was observed in the rerun; it does not establish that the price wording reliably fixes budget reasoning.
These counts concern the specific errors described, not an overall accuracy score. Two answers per version and model cannot establish a general error rate. The concrete result: the tested clarification removed the observed founding-price ambiguity in these GPT and Gemini reruns.
WHY THIS MATTERS
Find the wrong answer.
Test a precise fix.
A visibility score would not explain why the model invented a five-brand package. Reading the answers exposed a specific misunderstanding. Testing one source edit showed a better answer on the same model.
Next, we would separate the one-time and recurring budgets more clearly and run the check again. The pricing ambiguity improved. The remaining budget confusion gives us the next action.
OPEN THE EVIDENCE
Real calls.
Readable evidence.
OpenRouter prompts and answers ↗ · Results and paired checks ↗ · Method and failures ↗
A controlled source-pack experiment with one brand and no competing offers. No web search, organic visibility measurement, or sales measurement. A separate API harness ran the calls. Qwen returned eight unusable outputs; they are retained in the evidence. Full model names, prompts, settings, and limitations are included.
Earlier Gemini answers ↗ · Gemini clarification rerun ↗. The earlier Gemini 3.8 run was incomplete and is excluded from the paired comparison.
FIELD NOTES
Signal. Contact.
Trajectory.
Three illustrative routes from an external finding to work a team can actually approve and send. Outcomes depend on third-party editors, registries, and live retrieval.
01 / WRONG ENTITYSomeone else's record is attached to your name.
Find: a similarly named company and an outdated directory record appear beside your brand in AI answers.
Target: publisher corrections desk and directory data steward.
Move: publish a canonical identity page, assemble registry evidence, draft a correction and entity-split request, then recheck comparable answers.
02 / MISSING CITATIONThe sources AI cites do not mention you.
Find: a relevant trade publication or comparison page repeatedly appears for buyer questions while your verified facts are absent.
Target: section editor, benchmark owner, or contributor desk.
Move: prepare an evidence-backed article and concise pitch, seek editorial acceptance, then connect any published coverage back to your owned sources.
03 / GRAPH COLLISIONKnowledge records mix two different brands.
Find: a registry or knowledge-graph record merges attributes from unrelated entities.
Target: registry maintainer or knowledge-base curator.
Move: map the conflicting identifiers, submit documented corrections through the official channel, and monitor whether the corrected record is retrieved.
YOUR BRAND’S QUESTIONS
Find your next
better answer.
Setup & Launch costs €1,499 per brand, paid once. It includes setup, up to 100 articles total in month one, and an initial external source audit. The first month’s Growth articles count toward the 100. Growth is €499/month from month two, with prioritized external action briefs.
One brand, one language, one website. Client fact check and approval required. Prices exclude VAT. Images, video, paid placements, third-party fees, and custom integrations are extra.