CASE STUDY 9 · HALLUCINATION DETECTION
EdTech Platform in Hyderabad: Correcting Three Factual Errors That Were Costing Enterprise Sales Calls
A Hyderabad-based corporate training and EdTech platform was losing enterprise sales calls because ChatGPT and Gemini were stating three factual errors about their company — wrong founding year, wrong client count, and a certification they had never claimed. Correcting the AI's description through entity management and structured content took six weeks and corrected the AI responses across four platforms, while one external source update remained pending.
| Sector | EdTech / Corporate Training |
|---|---|
| Location | Hyderabad, India |
| Anonymised | Yes — client name withheld |
The Challenge
The company had been in business since 2019 and had grown to serve 80 corporate clients. In Q1 2026, their enterprise sales team started hearing a new objection on calls: "ChatGPT told us you were founded in 2015 and have over 500 clients." The company had been founded in 2019 and had 80 clients. The gap between the AI's description and the accurate facts was being used by some prospects as a reason for scepticism — if the basic facts were wrong, what else might be wrong?
Separately, Gemini was describing the platform as "ISO 9001 certified for training delivery" — a certification the company had never held or claimed. Two procurement teams had included this certification in their vendor evaluation criteria, which the company then failed to meet at the formal assessment stage.
What the Diagnostic Found
The "founded in 2015" error traced back to a 2021 blog post written by a freelance contributor that had incorrectly stated the company history. The post had since been updated but the original version was still in the Wayback Machine and was being read by AI training data sources.
The "500 clients" error appeared to come from an aggregator site that had published incorrect company data in 2023 and had not been updated.
The "ISO 9001 certified" hallucination had no traceable source — it appeared to be a generalisation error where AI associated the company's quality management language on their website with a certification that language implied but did not state.
No Wikidata entity existed for the company. No structured data on the website declared founding year, client count, or certifications explicitly. AI platforms had no authoritative machine-readable source for these facts.
What Dattva Did
Phase 1 — Source correction: the original 2021 blog post's incorrect version was identified in the Wayback Machine. The company requested removal via the Wayback Machine's exclusion process. The aggregator site was contacted directly with the correct company facts and a request to update. The company's own website was audited for any language that could be misread as implying ISO 9001 certification — all instances were either removed or explicitly stated as "not applicable."
Phase 2 — Entity creation: this work is Dattva's Hallucination Detection service — identifying and correcting inaccurate AI descriptions before they affect more sales conversations. A Wikidata entity was created for the company with verified facts: founding year 2019, headquarters Hyderabad, official website, current service categories. Every field was sourced from the company website and LinkedIn. A Crunchbase profile was completed with the same verified facts.
Phase 3 — Structured declaration: Organisation schema was added to the homepage with explicit fields for foundingDate (2019), numberOfEmployees, and areaServed. A "Company Overview" page was created with structured content explicitly stating founding year, current client count, and certifications held — written to be the definitive machine-readable source for company facts.
The Lesson
AI platforms fill gaps in their knowledge with confident-sounding generalisations. If a company's founding year, client count, and certifications are not stated explicitly in structured, machine-readable form on a verified source — Wikidata, Organisation schema, a factually sourced About page — AI will substitute what it infers. The inference is often wrong. The correction requires creating authoritative structured sources that give AI a confident answer before it has to guess.
See AI Hallucination Detection.
FAQ
How do factual errors enter AI responses about a company?
AI platforms build their knowledge of a company from whatever sources they can find — blog posts, press releases, aggregator sites, old news articles, Wayback Machine archives. An incorrect fact stated in any of these sources can propagate into AI responses. The error is then reinforced each time the AI finds the same fact in multiple sources, even if all of them trace back to the same original error.
Can a company force AI platforms to correct their description?
Not directly. AI platforms do not have a manual correction mechanism. The fix involves creating new, authoritative structured sources that give AI a more reliable answer than the incorrect existing sources. Wikidata is particularly effective because it is used by Claude and Gemini for entity verification. Organisation schema is effective because it gives Gemini and Perplexity a structured, citable source on the company's own domain.
What is the business impact of hallucinated certifications?
In B2B procurement, certifications are frequently included in formal RFP criteria. If AI tells an evaluator that a vendor holds a certification they do not hold, and that certification is included in the RFP requirements, the vendor will fail the assessment. The conversation at the assessment stage — "ChatGPT said you were ISO 9001 certified, but you have no evidence of this" — is extremely difficult to recover from, regardless of the actual product quality.
How long does it take for corrected facts to appear in AI responses?
Wikidata corrections typically propagate to Claude and Gemini within three to four weeks. Organisation schema corrections on the company website take four to six weeks depending on how quickly Bing and Brave re-crawl the updated pages. Aggregator site corrections depend on the site's own crawl and update cycle — some update within days, others take months. The total timeline for all platforms to reflect corrected facts is typically six to ten weeks.
See where your brand stands in AI answers today
Run a free AI Visibility diagnostic across ChatGPT, Perplexity, Gemini, and Claude — with prioritised, copy-paste fixes at no cost.