The 70% review threshold: the score below which AI stops recommending you

AI recommends. It does not rank.

Start with the distinction that governs everything else in this report. A search engine ranks. It orders ten blue links and lets you choose. A language model recommends. It synthesizes a verdict and hands you a short list, often three to five names, sometimes one. The mechanics are different, and so are the inputs.

When a model recommends, it is not retrieving the highest-authority page. It is assembling a consensus from the sources it trusts and producing a judgment about which brands belong in the answer. That judgment leans heavily on what other people say about you, not what you say about yourself. Social proof is the substrate. Reviews, forum threads, and third-party listings are the raw material the model reads to decide whether you are a safe thing to name.

This changes the unit of optimization. In search, you optimized a page to rank. In AI, you optimize a reputation to be recommended. The mention is the signal, and the strongest mentions come from places you do not own.

AI does not ask "who has the best page?" It asks "who do trusted third parties agree is worth naming?" Reviews are how that agreement gets recorded.

The data is now blunt enough to put a number on it. According to Onely, online reviews account for roughly 16% of the signals ChatGPT uses when deciding which brands to recommend. Authoritative list mentions carry the most weight at 41%, and awards and accreditations sit at 18%. Reviews are not the single largest factor, but they sit alongside list placement and recognition as one of three pillars of recommendation. Traditional SEO signals such as backlinks and keyword optimization carry near-zero direct influence on what gets recommended.

So the question is not whether reviews matter. It is whether you clear the bar.

The 70% floor

The most concrete finding in this space comes from SE Ranking, whose analysis of ChatGPT recommendation behavior across roughly 129,000 domains found that brands with an online review score below approximately 70% are significantly less likely to be recommended. Above the line, you are a candidate. Below it, you are quietly filtered out before the answer is even drafted.

Treat 70% as a floor, not a target. It is not the score that wins recommendations. It is the score below which you stop being eligible to win them. A model assembling a list of project management tools or local contractors will skip a brand whose aggregate sentiment reads as a liability, because the model is optimizing for a safe, defensible answer. A sub-70% brand introduces risk it has no reason to take.

This is why the threshold behaves like a gate rather than a dial. The relationship is not smoothly linear, where every additional point of review score buys a proportional increase in recommendation odds. Below the floor, the brand is largely absent. Above it, other signals such as list placement, recency, and platform breadth take over and decide who actually gets named.

The 70% line is the difference between being eligible and being invisible. It does not make you the recommendation. It makes you a candidate for it.

The practical implication is uncomfortable for brands carrying a backlog of unanswered complaints. A handful of one-star reviews left unaddressed does not just dent your average. Past the tipping point it can remove you from the answer set entirely, across every query in your category, for as long as the sentiment persists. Onely notes that AI's consensus-based approach means negative coverage from sources it trusts carries outsized weight, and a single authoritative negative can surface in recommendations for months.

There is a deeper reason the score behaves as a gate. A language model is not averaging your stars. It is reading reviews as evidence and constructing a narrative about you. When it scans a review profile it checks the aggregate score, identifies recurring themes in the feedback, and paraphrases what customers say into the answer it gives. A profile that sits below the threshold does not just lower a number. It hands the model a stock of negative themes to paraphrase, and the model will paraphrase them. The score is a proxy for the story, and below 70% the story stops being one a model wants to repeat.

This is also why the floor is consistent across categories rather than tuned to any single vertical. The model is making a risk calculation that does not change much whether the query is about accounting software or a roofing contractor. It wants to name brands it can defend, and a brand whose own customers are visibly unhappy is hard to defend. The 70% line is, in effect, the model's risk tolerance expressed as a number.

Having a profile beats almost everything

If the 70% score is the quality gate, presence is the entry gate, and the entry gate turns out to be the bigger story.

The most striking 2026 data comes from Trustpilot, whose "What AI says about you" report, conducted by Seer Interactive in March 2026, analyzed more than 800,000 AI responses across ChatGPT, Gemini, Perplexity, and Google AI Mode. The headline: businesses that actively gather and respond to customer reviews are cited in 75.3% of AI answers. Businesses with no active review profile are cited in just 1%.

Read that gap again. The difference between an absent brand and an engaged one is not a few percentage points. It is the difference between near-invisibility and being named in three out of four answers. The same study found a brand can lift its citation rate from 1% to 53.5% simply by establishing a review presence, then push it to 75.3% by collecting more than 80 reviews and responding to them regularly.

This reframes the work. Before you obsess over your score, make sure you exist on the platforms the model reads. Most brands' real problem is not a 68% rating. It is a profile that is thin, stale, or missing entirely.

Breadth compounds the effect. Multi-source analysis of AI citation behavior found that brands with active profiles on two or more review platforms such as G2, Capterra, Trustpilot, and TrustRadius are 3.4x more likely to be mentioned in ChatGPT responses than brands with none, and that presence across G2, Capterra, Trustpilot, and Yelp raises citation probability roughly 3x. The mechanism is consensus stacking. The model looks for agreement among trusted sources, and a brand that shows up favorably in several independent places reads as more verifiable than one that shows up in a single place or nowhere.

• Step one: claim and complete a profile on the review platforms that matter for your category.
• Step two: clear the 70% score floor.
• Step three: build to 80-plus reviews and respond to them.
• Step four: repeat across at least two platforms to trigger consensus stacking.

Where the citations actually go

Knowing reviews matter is not the same as knowing where the model looks. The 2026 citation data is specific, and it points away from your own website.

Omniscient Digital, analyzing more than 23,000 AI citations, found that for branded queries, 57% of citations go to reviews, listicles, forums, and case studies. Directory sites captured 17%, product pages 12%, and thought-leadership content only 5.4%. The interpretation matters as much as the number: branded queries ask the model to validate a decision, not to teach a concept. When someone asks "is Brand X any good?" the model reaches for social proof, not for your marketing.

Review and trust platforms have become a structural fixture of AI answers, not a niche source. The Trustpilot / Seer Interactive study found that review and trust websites are now the second most cited source type, accounting for 14% of all citations in AI responses, with Trustpilot the most frequently cited among review platforms.

Then there is Reddit, which behaves like its own category of social proof. The knowledgebase corpus and 2026 tracking both put it at the top of the citation stack. Digital Bloom found Reddit makes up 46.7% of Perplexity's top sources. Independent 2026 tracking found Reddit's citation share grew at least 73% in every category measured between October 2025 and January 2026, with roughly 24% of Perplexity's January 2026 citations and 44% of Google AI Overviews' social citations coming from Reddit. SE Ranking found that Reddit and Quora mentions correlate with roughly 4x higher citation likelihood in ChatGPT.

For branded queries, the model trusts the crowd over the brand. Reviews, forums, and listicles take 57% of citations. Your product pages take 12%.

There is nuance in how Reddit gets cited that is worth holding onto. 2026 tracking found that Reddit's raw citation frequency across all query types fell by roughly half, yet sole-source citations, the cases where the model cites Reddit and nothing else, rose 31%. The model is becoming more selective about when it reaches for Reddit, but when it does, it leans on it harder. For a brand, that means a single well-regarded thread can carry disproportionate weight, and a single damaging one can too.

The lesson is not to abandon your own content. It is to recognize that for the high-intent, late-funnel moment when a buyer is deciding between you and a competitor, the model is reading the rooms you do not control. Earning a place in those rooms is the work. The corollary is that you cannot fake it. Both Reddit communities and the models trained on them are tuned to detect promotional content, and inauthentic engagement damages your standing in both places at once. Social proof is only proof if it is real.

Volume, recency, and the 3Rs

A static review profile decays. AI systems privilege fresh signals, and reviews are no exception.

Trustpilot frames the mechanics as three factors, the 3Rs: relevance, recency, and ranking. Relevance comes from reviews containing detailed first-hand descriptions of products and services, exactly the specific language a model needs to answer a specific question. Recency comes from a steady flow of new reviews giving the model fresh material to draw from; Trustpilot reported its platform took in an average of 200,000 reviews a day in 2025. Ranking comes from the strong domain authority that keeps review pages surfacing in the first place.

Recency is not a soft preference. Onely found that 71% of ChatGPT citations come from content published in 2023 to 2025, and that products with 500-plus reviews in the last three months significantly outperform those with stale profiles. A review profile that peaked two years ago is, to the model, an old profile. The signal it sends is that the brand is no longer active, no longer current, no longer the safe pick.

Consumer behavior confirms why the model weights this the way it does. Yext's 2026 Consumer Search Behaviors research found that after an AI recommendation, star rating is the number one purchase influencer at 34%, followed by word of mouth at 30%, review recency at 29%, review sentiment at 28%, and review count at 28%. The model is approximating what the buyer will check next. It front-loads the same signals the human will use to validate the recommendation, because a recommendation the buyer immediately rejects is a bad recommendation.

That validation loop is real and measurable. Yext found that after getting an AI recommendation, 62% of consumers immediately search Google, 58% visit the business's website directly, and 52% click through to the sources cited in the AI response. Trust in AI does not eliminate verification. It only changes how people do it. Your review profile has to survive the second look, not just the first mention.

Being seen versus being selected

There is a trap in this data, and it is worth naming before you act on it.

SparkToro, in research with Gumshoe.ai spanning 2,961 prompts across ChatGPT, Claude, and Google AI Overviews, found that AI tools produce a different brand recommendation list more than 99% of the time when given the same prompt. The same list in the same order appears less than 0.1% of the time. Recommendation, at the individual-query level, is close to random.

This is not a reason to dismiss the work. It is a reason to do it correctly. Individual recommendations are volatile, but the probability of appearing is not. A brand that clears the 70% floor, carries a deep and recent review profile, and shows up across multiple platforms raises its odds of landing in any given list, even though no single list is guaranteed. You are not buying a slot. You are buying a higher draw from a deck that gets reshuffled every time.

There is also a dimension the SparkToro researchers explicitly did not measure: sentiment. They tracked whether a brand appeared, not how it was described. A brand named in 60% of responses but consistently hedged with caveats is in a very different position than one named less often but described with confidence. Being seen is not being selected. Appearing in the answer with a qualifier attached is closer to a warning than a recommendation. This is where review sentiment, not just review score, does its quiet work. The model reads the texture of what people say, not only the average star count, and it carries that texture into how it frames you.

So the discipline is straightforward. Stop optimizing for a single screenshot of a single answer. Optimize the underlying reputation that determines your probability of appearing favorably across thousands of answers. That is a social-proof problem, and it is one you can actually move.

Individual AI recommendations are nearly random. Your probability of appearing in them is not. Reviews are the lever that moves the probability.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.