You would not accept "our containers clear faster" from a supplier without asking faster than what, measured how, over how many shipments. You would ask automatically, because it is your industry and you know where the tricks are.
You have not built that instinct for this field yet. This chapter is the shortcut.
The seven tells of a manufactured statistic
- No sample size. "Studies show pages with X get 3× more citations." Which pages? How many? Over what period? A real study leads with its sample because it is the strongest thing about it.
- A suspiciously clean multiple. "2×", "3.2×", "10×". Real measurements are messy. Clean multiples are usually reverse-engineered from a conclusion someone already wanted.
- The seller benefits precisely. A chunking vendor publishing that chunking yields 2–4× improvement. Not automatically false — automatically requiring the method.
- Circular sourcing. Follow the citation. Blog A cites Blog B cites Blog C cites Blog A. This is extremely common in this field, and once you start following links you will find it constantly.
- A percentage with no baseline. "Improved citations by 292%" — from what to what? From one citation to four is technically 300%.
- Precision that cannot exist. "Each blocked bot costs 18–34% of potential citations on that engine." Nobody outside the platform could measure that. The range exists to look researched.
- Contradicts the platform's own documentation without acknowledging it. If a vendor's claim conflicts with Google's published guidance and they do not mention the conflict, they either have not read it or hope you have not.
The three-question test
Before you repeat any number from this field — including one from this book — ask:
Not disqualifying. Vendors run the best studies in this field because they have the data. But it sets your prior.
Sample size, time period, what was compared, control group. If it takes more than one click, it usually does not exist.
The strongest test. Given how retrieval works — Chapter 1 — does this finding make sense? A claim that contradicts the mechanism needs extraordinary evidence. A claim that fits it needs less.
Worked examples, using numbers from this book
"Pew found 8% click-through with an AI summary versus 15% without." Named non-commercial research organisation. Method published: 900 US adults, real browsing data, 68,879 queries, March 2025. Observed behaviour rather than a survey. Fits the mechanism — if the answer is on the page, fewer people click. Trust it.
"Branded web mentions correlate 0.664 with AI Overview visibility." Named vendor, 75,000 brands, stated method, full table published, and the authors themselves warn against reading causation into it. Fits the mechanism. Trust the direction; do not treat it as a dial.
"Content chunking produces 2–4× improvement in citation rates." No sample, no method, clean multiple, published by people selling chunking, and it contradicts Google's own documentation saying chunking is not required. Discard.
"98.8% of local businesses are invisible to AI." No method, alarming, precise to one decimal place, and it exists to create urgency. Discard.
A live example, from building this book
The worked examples in Chapters 9, 10 and 14 use real figures from DNE Logistics' operational records. Pulling them produced a textbook illustration of this chapter, so it is worth showing you the working.
The first query returned Port Klang import volumes for 2024 and 2025:
| Year | Import declarations | Containers |
|---|---|---|
| 2024 | 1,718 | 4,597 |
| 2025 | 3,928 | 10,400 |
| Apparent change | +128.6% | +126.2% |
A 126% year-on-year jump. It is a wonderful number. It is also completely false, and it would have gone into this book unchallenged if the next query had not been run.
Breaking the same data down by month showed that 2024 contains no records before August. The operational system was adopted mid-year. "2024" in that table is five months of data being compared against twelve. Volumes are, in reality, close to flat:
| Comparable window | Declarations | Containers |
|---|---|---|
| August–December 2024 | 1,718 | 4,597 |
| August–December 2025 | 1,745 | 4,704 |
| Real change | +1.6% | +2.3% |
+126% became +2.3%. Same database, same query language, same afternoon. The difference was one additional question: does my denominator cover the period I am claiming?
Nobody fabricated anything. A competent analyst ran a correct query against real data and got a wildly misleading answer, because a system-adoption date created a coverage gap that the totals concealed. This is how most bad statistics are actually born — not from dishonesty, but from a denominator nobody interrogated.
It is also why the Tier system exists. A vendor publishing "+126%" here would not be lying. They would simply not have run the second query, and you would have no way to tell from the outside. That is exactly why "what is your method?" is the only question that reliably works.
Apply it to this book
Everything here carries a tag and Appendix B lists the sources. Some of what I have rated Tier B will turn out to be wrong. The GEO paper's exact percentages are from 2024 systems that no longer exist, and I have said so in the chapter rather than in a footnote.
What I have tried to give you is not a list of facts to memorise but a mechanism to reason from, and the habit of asking where a number came from. The tactics in this book have a shelf life of perhaps eighteen months. The mechanism and the scepticism should last considerably longer.
The next time anyone quotes you an AI-visibility statistic, ask one question: "where is that from?" The answer will tell you everything you need to know about the rest of the meeting.