Four sources. Two judgments. My own score just got cut in half.

🚀 The AI Trend Oracle Report

The Compounding Math Nobody Checks

Recommendation is not won once. It accumulates. But accumulation only works if the things accumulating are actually separate, and this month gave us two expensive reminders that they often are not.

Last week I wrote that evidence about AI visibility has to survive the same scrutiny we tell clients their brand evidence must survive. I said it applies to vendors. I said it applies to me.

Then it did.

🚨 Market Intelligence

A Source Can Disappear in Six Days

On August 20, the story moving through this industry was that Reddit had nearly vanished from ChatGPT's citations.

The numbers, from the tracking firm Promptwatch: Reddit's share of ChatGPT Search citations fell from a 3.83 percent average between July 18 and August 7 to 0.52 percent between August 14 and 17. That is an 86.4 percent drop.

It happened in two stages. A first decline on August 8, then a sharper collapse on August 14.

The leading explanation is that ChatGPT changed its background search behavior on August 8, with site-specific queries rising from 0.37 percent to 16.8 percent of fanout queries. That is roughly a 46-fold increase, and it is a real change.

It also does not explain the second drop six days later. Nobody has accounted for that gap.

There is precedent for the uncertainty. A similar Reddit citation collapse in September 2025 was eventually attributed to Google removing its num=100 search parameter, which had nothing to do with Reddit or OpenAI at all.

Separately, on August 19 the IAB published Version 2 of its AI Transparency and Disclosure Framework. Different market, different problem, and one principle in it stopped me: not every use of AI needs a label, because labeling everything teaches people to ignore labels. Hold onto that. It comes back later in this issue in a form I did not expect.

The thing to take from this is narrower than most of the commentary around it.

Any single source can lose most of its weight in under a week, for reasons that are not visible from outside and may not be about that source at all.

What this data cannot tell you is what happened to any particular brand. A category-wide citation share is an aggregate. If you were relying on Reddit as an evidence source, this is a reason to go and check your own position. It is not a finding about your position. Half the posts I read last week made that jump, and Promptwatch had already said their own number was provisional.

📊 What's Actually Happening

A Measurement Vendor Did the Right Thing, and It Barely Got Noticed

Here is the part of that story I have not seen anyone highlight.

Promptwatch published a dramatic, highly shareable number. An 86 percent collapse. It traveled everywhere, including into a hundred LinkedIn posts declaring Reddit strategy dead.

And Promptwatch said, in their own reporting, that the size of the drop is **provisional**, and that they cannot yet rule out a data-collection issue on their end.

Sit with how unusual that is.

They had the most attention-grabbing figure in the category that week. They attached a caveat to it that undercuts their own headline. They named a possible fault in their own instrument rather than waiting for someone else to find it.

Three weeks ago the IAB published Measuring Visibility in the AI Era and defined decision-grade measurement as requiring reproducibility, disclosed variation ranges, sample sufficiency and methodology transparency. Most of us read that as a standard someone would eventually have to enforce.

Promptwatch just did it voluntarily, on a number that would have been more useful to them uncaveated.

That is what the standard looks like when it is working. Not a certification body. A vendor volunteering the thing that weakens their own claim, because a claim that survives that is worth more than one that avoids it.

I want to be clear that this is a compliment I am paying to a company I have no relationship with, about a number I cannot independently verify. Which is rather the point.

🔍 Featured Insight

I Ran a Scan on Myself and Found the Flaw

On Friday I ran an Axis Suite scan on Axis Suite.

It came back better than the week before. Software Advice had started citing us. The panel read: **4 independent sources. G2, Capterra, GetApp, Software Advice.** Evidence score 40 out of 100.

I was pleased for about an hour.

Then I looked at the actual quote the engine had returned:

> "5.0/5 on Software Advice, G2, and GetApp"

The same rating. Across three sites. That is not three parties reaching the same conclusion. That is one set of reviews rendered three times.

So I went and looked. What I can tell you firsthand: I applied to each of those properties separately, each approved separately, and they are administered through one shared dashboard. What I observed in the scan is that the identical rating surfaced across all three. I know the duplication happened in my case because I received one review, not three. That one review appeared across Capterra, GetApp, and Software Advice. Third-party reporting also describes the three properties as sharing a review catalog. So whatever value those three domains have as separate retrieval surfaces, they were not three independent judgments about Axis Suite. They were one judgment appearing in three places.

I am telling you what I saw and what is reported rather than claiming I have audited their backend, because that distinction is the entire subject of this issue.

There is more. G2 acquired Capterra, Software Advice and GetApp from Gartner. Announced in January, now closed. The catalogs are still separate today, so a G2 review genuinely is a different judgment. But one corporate decision away, it is not.

So the honest count was never four.

**4 sources named us. 2 independent judgments corroborated us.**

Evidence score after the fix: **20 out of 100.** Half of what my own product told me on Friday.

And there was a second one in the same scan. Stanford's AI Index appeared as an independent source. It had been cited in an answer about LLM citation bias. It said nothing whatsoever about Axis Suite. The system counted a source that appeared in the answer as a source that corroborated the brand.

Two failures, one shape:

**A citation is not evidence. A repeated citation is not repeated evidence.**

That is the exact error I have spent months telling this market to look for in other people's tools. It was in mine. I found it by scanning myself and reading the quote instead of the number.

I am publishing the number going down because last week I wrote that if Axis Suite publishes a finding, the method should travel with it, and that otherwise I am asking the market to take my word for it. A methodology correction that lowers my own score is the only version of that sentence worth anything.

🎯 This Week's Strategic Lens

Correlated Sources Do Not Compound. They Echo.

I promised last week to look at recommendation as something that accumulates rather than something you win. Here is the version I did not expect to be writing.

The compounding argument goes like this. The first source that says something about you establishes a claim. The second makes it harder to dismiss. Somewhere past that, the claim stops being something a model has to weigh and starts being something it treats as settled.

I teased this last week as the fifth or sixth signal being worth more than the first four combined. I am walking that specific phrasing back, because I cannot support that ratio and I spent last week's issue objecting to numbers nobody can support. The direction is right. The arithmetic was mine and it was decoration.

What I can defend is the condition underneath it, and it is the part almost nobody checks.

**The sources have to be independent of each other.**

If your five signals are five renderings of one underlying source, you do not have compounding. You have an echo.

I cannot tell you what that does inside a model, and neither can anyone else selling you a tool. What I can tell you is what it does to the measurement, because that part is observable: repeated versions of one judgment are indistinguishable from independent corroboration unless somebody resolves the relationship between the sources. Your report cannot see the difference. Nor could mine, until Friday.

Which is the Reddit story, restated. Brands with concentrated evidence in one source did not have a diversified position that happened to include Reddit. They had one source producing many citations, and when it moved, the whole position moved.

So there are two different numbers, and most reporting fuses them:

**Retrieval sources.** How many distinct places an engine can pull you from. Four separate domains means four sets of authority signals and four chances to land in an answer. Real, and genuinely better than one.

**Independent judgments.** How many separate parties actually reached a conclusion about you. Three storefronts on one catalog is one judgment.

A citation tracker can only ever report the first. The second is what tells you whether your position survives one source having a bad month.

This is where that IAB line comes back. Labeling everything teaches people to ignore labels. Counting everything as independent corroboration teaches you to ignore the count, and it took my own product doing it to me before I saw that they are the same failure. A marker only carries information if it is sometimes withheld.

I am not adding a fifth P and I am not naming a new framework. It is one question to ask of any evidence panel you are shown, including one of mine:

**How many of these would still be there if one of them went away?**

👨‍💻 Founder's Note

I have been noticing something about how this month has gone, and it is not comfortable.

Every genuinely useful thing I have learned in the last three weeks came from something being wrong. A scan that returned a confident zero for a company that was simply filed in the wrong category. A comment thread where a stranger corrected my statistics and was right. My own Evidence score, which I had been quietly pleased with, resting on a count that was double what it should have been.

I used to think building an intelligence product meant getting more things right than other people. I do not think that anymore. Everybody in this category is wrong about something, constantly, because the ground moves weekly. What separates products is not the error rate. It is whether the thing tells you when it does not know, and whether the person building it goes looking.

The version of Friday where I do not read the quote and just look at the 40 is available to me every week. It is faster and it feels better. It is also how you end up shipping a number that somebody eventually discovers is wrong, at the worst possible moment, in front of people whose opinion of them depends on it.

I would rather find it on a Friday afternoon scanning myself.

There is one real advantage to being at the stage I am at, and it is worth naming rather than talking around. Finding this now cost me an afternoon and a paragraph. Finding it in two years, with that number sitting inside other people's quarterly reviews, would have cost something I could not pay back.

Every product in this category is going to have a version of my Friday. The only variable is how much is riding on the number by the time it shows up.

- Dana Billingsley | Founder, Axis Suite

🛠 THE 10-MINUTE INDEPENDENCE AUDIT

Pull up whatever list of sources your AI visibility tool says is corroborating you. Run it against five questions. This works on any vendor's report, including mine.

**1. Who owns each one?** Look up the parent company of every source on the list. If two share a parent, flag them. This takes about four minutes and almost nobody does it.

**2. Do any of them share content?** Ownership is not the real test. Shared content is. Two sites under one parent with separate editorial are two judgments. Three sites running one review catalog are one. Check whether a review or listing submitted to one appears on the others.

**3. What is each source actually saying about you?** Not "did it appear." What claim is it making, and does that claim name you? A source cited in an answer to support a point about the category is not evidence about your brand. Read the quoted sentence, not the source name.

**4. If your largest source disappeared tomorrow, what remains?** Cover it with your thumb. Count what is left. That number is your real position, because that is the scenario Reddit users just lived through with six days of warning and no explanation.

**5. Is the count of sources being reported as the count of corroborations?** These are different measurements. A good report shows both. Most show one and calls it independence.

If four of those come back clean, you have much stronger evidence that your position is genuinely diversified. If they do not, you now know where the concentration is, which may be worth more than the score itself.

☕AI Caffeinated Wisdom

Your Weekly Dose of Caffeinated Wisdom

You hear a piece of news from someone on Tuesday. You are mildly interested.

You hear it again Wednesday from someone unconnected to the first person. Now you are paying attention.

By Friday, four different people have told you the same thing, and you have stopped questioning it entirely. You start repeating it yourself. You make a small decision based on it.

Then, weeks later, it turns out all four of them heard it from Linda.

Nothing about your reasoning was faulty. Four independent confirmations genuinely should raise your confidence, and if the four had been independent, you would have been right to believe it. The failure was not in the arithmetic. It was in a premise nobody checks, because checking it requires asking each person where they heard it, which nobody does, because it sounds like you are calling them a liar.

This is the whole thing.

Confidence built on repetition is only as strong as the independence underneath the repetition. And independence is invisible by design. Four people telling you the same story looks exactly like four people telling you the same story, whether it came from four places or one.

A measurement system sees your brand named on four sites and can count four signals unless it knows those signals share the same origin.
The question worth asking, of your evidence and of Linda's, is not how many times you heard it.

It is how many places it actually came from.

Stay steady. ☕

🔔 CLOSING SIGNAL

Recommendation compounds. That part is real and it is the most valuable thing about this whole category.

But it compounds across independent sources, and independence is the one property that never shows up on a dashboard, because it is not a property of any source on its own. It only exists in the relationships between them.

This month gave us both halves. Reddit showed what happens when a concentrated position meets a change nobody can explain. My own scan showed what happens when four sources turn out to be two.

My bet is that the brands most likely to hold their position through the next model update will not simply be the ones with the highest score this week. They will be the ones whose evidence comes from enough genuinely separate places that no single change can take it all.

Count the sources. Then count where they came from. The second number is the one that holds.

🛠 Coming Next Week

I am going to be honest about where this one stands, because it is not written yet and I do not yet know what it says.

Last week I got into a long debate on Reddit with several people who build in this space, about whether the diagnostic tables everyone is designing actually work. It came down to a question none of us could answer: when you ask an AI engine the identical question twenty times, does it even perform the same kind of task each time? Does it explain the category on some runs and produce a shortlist on others?

Every cross-engine comparison being proposed right now assumes it does not move. Nobody has measured it.

So this week I am going to measure it. Three prompts, two engines, twenty runs each, one hundred and twenty responses, hand-labeled blind after the fact so I cannot bend the result toward what I expect.

The method is already written down and posted publicly, including the threshold and what I will conclude if the answer turns out to be boring. That happened before any runs, which is the only order in which writing it down means anything.

Next week I will publish what came back, whichever way it lands. If the answer is that engines are perfectly consistent, then a lot of what I have argued this month needs revising and I will say so.

That is the whole point of writing the method down first.

🚀🤖✨📊🎨

The Axis Suite

AI Recommendation Intelligence. AI Narrative Defense. Agentic Visibility Infrastructure.

"Axis Suite is the independent intelligence layer that explains what AI believes about your brand, why it believes it, and what decision that belief ultimately drives."

👉 Axis Suite | Read the Proof Center

📬 Thank you for being part of this.

Here's to staying visible, staying understood, and becoming the obvious choice.

Stay caffeinated, stay inspired. See you next week! ☕

From the trenches,
The Axis Suite Team 💪