Google AI Overviews is 91% accurate according to a New York Times/Oumi study of 4,326 searches — a six-point improvement from 85% in October 2025. But at 5 trillion searches per year, the 9% error rate means tens of millions of wrong brand descriptions every hour. And 56% of correct answers can't be verified through the linked sources.
That 91% sounds safe. It isn't.
The New York Times, working with AI startup Oumi, tested those 4,326 Google searches using SimpleQA — a factual accuracy benchmark originally developed by OpenAI. They ran the tests in two rounds: once in October 2025 with Gemini 2 powering AI Overviews, and again in February 2026 after Google upgraded to Gemini 3.
Google calls this progress. They're right — it is. But the maths at Google's scale tell a different story.
Google handles over 5 trillion searches per year. At a 9% error rate, that's tens of millions of wrong answers every single hour. Not every hour of a bad day. Every hour. All the time.
If your brand shows up in any of those wrong answers — with incorrect pricing, outdated product claims, or a competitor's feature attributed to you — that misinformation reaches millions of people before anyone at your company even notices.
The source verification problem is worse than the accuracy problem
Here's the finding that should worry brand marketers more than the 9% error rate.
Of the answers AI Overviews got correct in February, 56% were "ungrounded" — meaning the linked sources didn't actually support the answer. With Gemini 2 in October, that number was 37%. It got worse as accuracy improved.
Think about what this means: AI Overviews is increasingly confident in its answers but increasingly disconnected from the sources it cites. It's getting better at sounding right while making it harder for anyone to verify whether it is right.
For brands, this creates a specific problem. When AI Overviews describes your product, your pricing, or your positioning — and the linked source doesn't say what AI claims it says — there's no way for the reader to fact-check. They take the AI answer at face value. And they should — it's Google.
Facebook ranked as the second most-cited source (5% of correct answers, 7% of incorrect ones). Reddit placed fourth. These aren't authoritative brand sources. These are platforms where anyone can say anything about your product.
This is a Layer 3 problem in the AI Visibility Stack
Most brands think about AI visibility as a binary: either AI mentions you or it doesn't. That's Layer 1 (Discoverability) — can AI find your content?
But the AI Overviews accuracy data reveals a much harder problem at Layer 3 (Authority): when AI does mention you, is it getting the story right?
The 3-Layer AI Visibility Stack breaks down like this:
| Layer | Name | Key Question | Fix Horizon | Focus Area |
|---|---|---|---|---|
| Layer 1 | Discoverability | Can AI crawlers find your content? | Days to weeks | robots.txt, schema markup, llms.txt |
| Layer 2 | Citability | When AI finds you, does it cite you? | Weeks to months | Content structure, definitive language, original data |
| Layer 3 | Authority | When AI cites you, does it position you correctly? | Months to quarters | Narrative control, third-party coverage, share of voice |
The 91% accuracy finding hits Layer 3 directly. Your brand may be discoverable (Layer 1) and citable (Layer 2), but if AI Overviews describes you inaccurately 9% of the time — and can't even verify its correct descriptions 56% of the time — you have an Authority problem that no amount of technical SEO will fix.
What the error patterns look like in practice
Want to know how your brand scores on these same metrics?
We'll run 20 prompts across 3 AI platforms and send your report within 24 hours.
The study documented specific error types:
- Wrong dates: AI Overviews said Bob Marley's museum was established in 1987. The correct year is 1986. Off by one year, but confidently wrong.
- Contradictory citations: For Yo-Yo Ma's Classical Music Hall of Fame induction, AI linked to the correct source but then claimed "no record" existed.
- Correct context, wrong detail: Dick Drago's age at death was correct, but the date of death was wrong.
Now imagine these error patterns applied to your brand:
- "Brand X's flagship product starts at ₹12,999" when it actually starts at ₹9,999
- "Brand X was founded in 2019" when it was founded in 2018
- AI links to your pricing page but states a different price than what's on it
These aren't hypotheticals. If AI Overviews makes these errors on verifiable historical facts with one correct answer, it will make them on brand information — where multiple sources often provide conflicting data.
Why this matters more for Indian brands
India has one of the world's largest Google user bases — with over 97% search market share. AI Overviews has been rolling out progressively in India throughout 2025-2026, and now appears in 14% of shopping queries globally.
Indian D2C brands face a compounding problem:
-
Less authoritative English-language coverage. Many Indian brands have thinner Wikipedia pages, fewer international press mentions, and less structured third-party data than their Western counterparts. AI has fewer reliable sources to draw from — which means it's more likely to get things wrong.
-
Rupee-dollar conversion confusion. AI regularly confuses INR and USD pricing, or cites outdated pricing from old listicles. We've seen this in our own Cited Index data across 257 Indian brands — pricing accuracy in AI responses is one of the weakest signals.
-
Category-level misinformation. When AI gets a category description wrong — "Mokobara is a budget luggage brand" vs. "Mokobara is a premium travel brand" — that framing shapes every subsequent recommendation. One wrong description cascades across thousands of queries.
Three things to do this week
You can't fix Google's 9% error rate. But you can reduce the odds that your brand is in that 9%.
1. Audit what AI actually says about you right now
Run your top 10 category prompts through Google AI Overviews, ChatGPT, and Perplexity. For each response, check:
- Is the brand description accurate?
- Is the pricing correct and current?
- Are the product features attributed correctly?
- Is the competitive positioning fair?
Document every error. This is your Layer 3 baseline — Brand Sentiment (Metric 4 in the Cited 8) and AI Citation Rate (Metric 1).
2. Control your source material
The 56% ungrounded finding means AI is often not pulling from your official pages. Fix this:
- Ensure your llms.txt file is complete and accurate
- Update your About page, product pages, and FAQ with definitive, extractable statements
- Add structured data (Product, Organization, FAQ schema) to every key page
- Publish a clear, authoritative "About" narrative that AI can extract verbatim
If AI doesn't have a reliable source from you, it'll pull from Facebook, Reddit, and outdated listicles instead.
3. Monitor continuously, not once
The study tested two snapshots — October 2025 and February 2026. Accuracy changed by 6 points in four months. AI Overviews is a moving target. What it says about your brand today may differ from what it says next month.
Set up monthly tracking using the Cited 8 metrics:
- AI Citation Rate (Metric 1): Are you being mentioned?
- Brand Sentiment (Metric 4): When mentioned, is the description accurate?
- Schema & Technical Health (Metric 7): Can AI crawlers access your authoritative content?
- Content Freshness (Metric 8): Is your content current enough to be cited?
Or use Cited's free GEO Score to check your site's AI-readiness in 30 seconds — it scans 15 signals across Discoverability, Citability, and Authority.
The headline is "91% accurate." The story is "56% unverifiable."
Google's response to the study was predictable: they called SimpleQA a "flawed benchmark" that "doesn't reflect actual search behavior." Fair enough — no benchmark perfectly mirrors real usage.
But the source verification finding isn't about the benchmark. It's about what happens after AI generates an answer. If 56% of correct answers can't be verified through the linked sources, that means AI Overviews is increasingly operating on its own reasoning rather than citing evidence. It's getting better at giving correct answers while getting worse at showing its work.
For brands, this is the real risk. You can't rely on Google's linked sources to represent you accurately. You have to ensure your own authoritative content is so clear, so structured, and so prominent that AI has no choice but to use it.
The 9% error rate is a scale problem. The 56% unverifiability rate is a trust problem. Both are your brand's problem now.
FAQ
Is AI Overviews getting more or less accurate?
More accurate — 85% in October 2025 to 91% in February 2026, according to the NYT/Oumi study. However, source verifiability declined from 63% to 44% over the same period. Accuracy is improving while traceability is degrading.
How many searches does Google process per year?
Over 5 trillion. At a 9% AI Overviews error rate, that translates to tens of millions of potentially incorrect answers every hour.
What is SimpleQA?
SimpleQA is a factual accuracy benchmark developed by OpenAI. It consists of over 4,000 questions with verifiable answers, used to test how accurately AI systems respond to factual queries. Google disputes its relevance to real search behavior.
What does "ungrounded" mean in this context?
An ungrounded answer is one where the AI's response is correct, but the linked source doesn't actually support the claim. The reader sees a citation that appears to verify the answer, but clicking through reveals the source says something different — or doesn't address the topic at all.
How can Indian brands protect themselves from AI misinformation?
Start by auditing what AI currently says about your brand across ChatGPT, Gemini, Perplexity, and Google AI Overviews. Control your source material with structured data, llms.txt, and definitive product descriptions. Monitor monthly using the Cited 8 metrics — particularly Brand Sentiment (Metric 4) and Schema & Technical Health (Metric 7).
Does this affect AI platforms beyond Google?
The study specifically tested Google AI Overviews. However, all AI platforms face similar accuracy challenges. ChatGPT, Perplexity, and Claude each have their own error profiles — which is why tracking across multiple platforms with tools like the Cited Index gives a more complete picture.