AI Overviews: your brand cited without search, the silent backside of negative reviews
Summarize this article with AI
« I lost 22% of traffic in 6 weeks and Google sent me no warning. »
A client calls me on a Tuesday morning. E-commerce, 4,200 SKUs, 47,000 organic sessions per month. For six weeks, qualified traffic is hemorrhaging. No manual penalty. No sudden drop in core keyword rankings.
He checks Search Console. Nothing.
He checks backlinks. Nothing.
He checks content. No technical issues.
Yet conversion rate drops from 3.1% to 2.4%. Monthly organic revenue loses €3,400.
I launch my audit.
I look at what AIs are saying about his brand.
The answer chills me.
Across four LLM comparators — Google AI Overviews, ChatGPT, Perplexity, Bing Copilot beta — his brand appears in answers to questions like « which CRM should I choose » or « reliable dropshipping platform ». And in these syntheses, the AI systematically cites two negative reviews posted 18 and 21 months ago. Reviews buried on Trustpilot that nobody read anymore.
Potential customers weren’t typing « [His brand] scam ». They were just asking for tool comparisons. And the AI served them his negative review. No filter. No context. No way for him to respond.
The sneaky mechanism: when AI cites you without being searched for
AI Overviews doesn’t work like classic search.
I observe with my clients that the trigger is no longer an intentional reputation query. It’s a comparison query. « What’s the best CRM? », « which platform to choose in 2026 », « what alternatives to X exist ».
According to Search Engine Journal, these tools actively scan forums, verified reviews, Reddit threads, social content to build comparative synthesis. They don’t distinguish between a representative customer review and an isolated complaint. They aggregate and weight by signal density. If your brand has three very detailed negative reviews and only twenty generic positive ones, the LLM can latch onto the contrast as « distrust ».
A Fast Company report, cited by Search Engine Journal, notes that AIs sometimes rephrase statements. A review that said « support took 48 hours to respond that day » becomes « customer service is slow » in the synthesis. Nuance evaporates.
It’s not a volume question.
It’s a question of structural toxicity of the signal.
A review from 18 months ago, 12 appearances in AI comparatives in 30 days
The e-commerce client I mentioned manufactures components. His market is technical B2B. His buyers search « industrial USB-C connector supplier comparison ». Not glamorous. But that’s his business.
I identified that two Reddit discussion threads and one 1-star Trustpilot review from 21 months ago formed the negative cluster. The review mentioned a delivery defect for a specific order, resolved since. But the AI doesn’t read review updates. It reads the initial timestamp and the rating.
Results after 30 days of observation:
- The negative review appeared in 12 distinct AI comparisons (Google, ChatGPT, Perplexity).
- Click-through rate on category pages concerned dropped 27%.
- Average time on site fell from 1 min 42 s to 48 seconds.
- Organic revenue for that product line fell by €3,400 net per month.
And all that for a review that Google classic placed on page 7.
The AIs put it on the front line, without cost to anyone.
Without counterpower.
Without recourse.
Why responding or deleting isn’t enough — the positive layer is the ultimate weapon
The natural reflex is to want to delete the review.
Report to the platform. Request removal. Fight.
I’ve helped more than 30 clients try this path. Fewer than 20% of requests succeed in less than 90 days.
And crucially: deleting a review doesn’t erase it from AI training data. The AI may have memorized the signal before deletion. And it will keep citing other negative sources if the ratio doesn’t change.
The real lever is to build a layer of positive content so dense and diversified that the AI prefers to cite that. Search Engine Journal explicitly recommends this: create case studies, video testimonials hosted on your domain, in-depth articles on comparatives, structured FAQ pages. Basically, proprietary signals with strong semantic authority.
The 4-step system I’ve built so your reputation no longer depends on LLM goodwill
I apply here the DOSE framework — Define, Observe, Structure, Engage — taught by Guillaume Attias at BMO Academy. This isn’t a checklist. It’s a permanent engine.
Step 1 – Audit of negative footprint
I identify all signals that LLMs cite. I use Google AI Overviews, ChatGPT, Perplexity, Bing Copilot. I run 40 sector comparison queries and note every mention of your brand. Result: a map of toxic clusters and recurring themes.
Step 2 – Prioritization by surface probability
A negative review on a low-traffic forum doesn’t carry the same weight as an indexed Reddit thread in Google Discover. I prioritize by volume of actual citations in AIs, not by star count. Often, 3 complaints generate 80% of the damage.
Step 3 – Respond and occupy the space
On third-party platforms, I write calibrated responses: contextual, factual, with a link to resolution. No pathos. In parallel, I push proprietary content: case studies, testimonials, comparatives, FAQ. Every piece is structured into semantic cocoons exactly as I do with classic SEO — except the target here is the AI.
Step 4 – Monitoring loop
I set up a surveillance system that scans the 40 key queries weekly for AI responses. The moment a new negative mention emerges, an alert fires. The client can act in 48 hours, not 6 weeks.
This system runs without me. That’s my obsession. One dashboard, alerts, documented procedure. The client doesn’t call me because he discovered the problem. He calls me because he wants to amplify his positive layer.

