Statistics and Data in AI Buyer Guides: The Numbers You Can Use, the Ones Generators Invent, and How to Cite Them
Numbers make a guide look researched, which is exactly why generated guides are full of them: "73% of commuters ride in the dark", "the global bike light market reached $2.1 billion", "a recent survey found". None of those came from anywhere.
The reader who checks one finds nothing and discounts everything else on the page, including the parts that were right. Meanwhile the numbers that would actually help (4,100 reviews at 4.4 stars, 90 minutes on full, the clamp complaint in one review in twenty) go unused because the tool didn't know they were the point.
This post covers which numbers earn a place in a buyer guide and where they come from, the three statistics generators invent and how to recognise them, how to source a real number in two minutes, the citation format that works on a product page, and why three real numbers beat twelve fake ones.
It's the numbers chapter of the hallucinations post, which covers the other things generators make up.
The short answer
Use numbers from the listing, the spec sheet and the review pattern: review counts and averages, price bands, compared specs, how common a complaint is. Delete any percentage, market size or survey that has no source you can link. Cite inline with the source and a date. Three real numbers in a guide do the work; the rest is decoration that can be caught.
The numbers that earn their place
| Number | Where it comes from | What it does for the reader | Example |
|---|---|---|---|
| Review count and average | The listing, through the API | Signals how many people have bought it and how they felt | "4,100 reviews at 4.4 stars as of October 2026" |
| Price band and movement | The API; price history where you track it | Whether now is a normal time to buy | "Usually $55 to $65; it dipped to $45 on Prime Day" |
| Compared specs | The spec sheets of the picks | The decision itself | "4 hours on medium vs the Lezyne's 2.5" |
| Complaint frequency | Your read of the one- and three-star reviews | How seriously to take the con | "The clamp complaint appears in roughly one review in twenty" |
| Return pattern | Reviews that mention returning it; your own data if you have it | Whether to buy with confidence | "Few reviewers mention returning it; the sizing complaints are about the strap version" |
| Your own test result | Only if you ran it | Real experience | "On my commute it lasted four nights on medium" |
Every row is a number about this product or this decision. That's the test: a statistic about the category or the world rarely helps anyone choose a bike light. The one-star reviews post covers the complaint-frequency row, which is the one most guides skip and most readers value.
The three generators always invent
- The tidy percentage with no source. "68% of cyclists say visibility is their top concern." The giveaway is a precise number and the words "studies show", "research suggests" or "experts estimate" standing in for a citation. Search the sentence; nothing comes back.
- The market size. "The global bike light market was valued at $2.1 billion in 2024." Sometimes a real report exists with a different number; usually not. Either way the reader choosing a light doesn't care.
- The survey nobody ran. "In a recent survey of 500 commuters…" If you didn't run it and can't link it, it didn't happen.
A fourth, subtler one: a real number attached to the wrong thing, like the parent listing's review count quoted for one variant, or the 1,100-lumen model's run time in the 800's section. The fact-check post covers checking numbers against the listing; this is the step before it, deciding which numbers should be there at all.
Sourcing a real one in two minutes
- Review count and rating: the listing, read through the API at write time, with the date.
- Price band: the API's current price plus, if you track it, the range over the last few months.
- Run time, weight, capacity: the manufacturer's spec page, not the Amazon bullets, which are often the top variant's.
- Complaint frequency: sort to one-star, read twenty, count how many mention the thing; "about one in five of the critical reviews" is honest and checkable.
- Anything about the category or the market: find the primary source (the trade body, the government statistic, the named study) and link it, or cut the sentence. "Cut" is the right answer nine times in ten.
The citation format for product pages
Inline, in the sentence, with a link where there is one and a date where the number moves: "Amazon lists 4,100 reviews at 4.4 stars (October 2026)"; "the manufacturer rates it at 4 hours on medium"; "across the twenty most recent critical reviews, four mention the clamp".
Footnotes and a references section are for essays; on a page that's scanned on a phone, the source has to sit next to the number or it isn't seen. A number with "according to" beside it reads as checked, and a number without reads as invented, whether or not it was.
Why three real numbers beat twelve fake ones
A reader forms a view of the page's reliability from the first number they check. A guide with three cited, product-specific numbers passes that check and the reader extends the trust to the verdict. A guide with twelve, of which nine are decorative and three are invented, fails it at the first invented one, and the reader discounts the verdict along with the statistics.
The same holds for the systems that lift passages into snippets and overviews: they cite specific, sourced claims and skip the ones that read as filler. Fewer numbers, each one real, each one about the product, is the whole rule.
Doing it with AffSERP
AffSERP's X Best Review tool draws every number in a guide from the listing data through Amazon's API (review count, rating, price band, the specs it compares) and from the review pattern it reads for the caveats, and its defaults ban the unsourced percentage, the market figure and the invented survey, so the three inventions above don't get written.
The date goes on the review count and price automatically. Your own test result, where you have one, is the one number it waits for you to add.
FAQ
Does AI make up statistics?
Routinely, and in a recognisable shape: a tidy percentage with no source ('studies show 73% of cyclists…'), a market-size figure ('the bike light market is worth $2.1 billion'), or a survey nobody ran ('according to a recent survey'). They're pattern completions: the genre contains such sentences, so the model produces one. None of them survive a search for the source, and a reader who checks one stops trusting the page.
What statistics should a product review include?
The ones the page can stand behind: the product's review count and average, its price band and whether it's moved, specs compared with the alternatives (run time, weight, capacity), how common a complaint is in the reviews, and a return-rate pattern where you can see one. These are numbers from the listing and the review pattern, and they're worth more than any market statistic because they bear on the decision.
How do you cite a statistic in a blog post?
Name the source in the sentence and link it: 'Amazon lists 4,100 reviews at 4.4 stars (as of October 2026)', or 'the manufacturer's spec sheet rates the battery at 4 hours on medium'. On a product page, the inline source beats a footnote: readers scan, and a number with 'according to' next to it reads as checked. Add the date for anything that changes.
Are statistics good for SEO?
A real, cited number is the kind of specific claim that gets quoted, lifted into snippets and cited in AI Overviews, and linked by other sites. An invented one is the kind that gets a page flagged as unreliable when a reader or a competitor checks it. The SEO value is entirely in the sourcing; the number itself does nothing.
AffSERP pulls review counts, ratings, prices and specs through the API, reads complaint frequency from the reviews, and bans the unsourced percentage and the market figure by default.
Start with free wordsKeep reading
