99.99%.

That's the brand-safety score Integral Ad Science slaps on its work on X. The platform formerly known as Twitter.

The platform currently known as the place where a chatbot named Grok occasionally free-associates about genocide before lunch.

Ninety-nine point nine nine percent safe. Suitability north of 97%. Independently measured, they say. On X. In 2026. Sure.

Here's the thing. I'm not reading that number off a slide in some windowless conference room in Cannes while a guy in unstructured linen says the word "trust" forty times. I'm reading it as a person who is actually on the platform they're grading. So let me tell you how safe X is. From the inside.

I'm an Orthodox Jew. I've got tens of thousands of followers on X. By every metric IAS claims to worship, I'm the good part of the neighborhood — the established, high-reach, brand-adjacent account that's supposed to be the safe, sunlit, 99.99%. And I get threatened and attacked almost every single time I post. Not when I say something spicy. Not when I pick a fight. Just... existing. Visibly Jewish, on the timeline, on a Tuesday. The trigger isn't what I say. The trigger is who I am.

So I barely use it anymore. Read that again, because it's the part that should make a media buyer spit out their cold brew. I built tens of thousands of followers and I walked away from the whole thing, because the daily cost of posting while Jewish got higher than the reach was worth. That's a real publisher making a real economic decision because of a real river of garbage that IAS's machine apparently cannot see. Because IAS's machine is over here, beaming, holding up a little sign that says 99.99% safe.

Line those two things up. A verification company is certifying, to Fortune 500 CMOs, that the environment I personally got chased out of is 99.99% brand safe. One of those two things is describing the actual X. Spoiler: it's not the one with the four nines.

And look — my timeline is not the story. My timeline is the tell. The story is bigger and dumber and more expensive than one guy's mentions.

The story is that the ad industry built an elaborate, gleaming, AI-powered apparatus whose entire job is to "protect" brands, and the apparatus does... approximately nothing. And I want to be precise about the nothing, because it is a specific and honestly kind of beautiful nothing. It's not that these tools sometimes miss. It's that when you actually go looking, the thing you paid to stand guard was asleep, or in the wrong building, or — more likely — never really a guard at all. It was a dashboard with a nice green number on it. And the nice green number was the product.

This is a company whose literal, printed-on-every-press-release mission is to be "the global benchmark for trust and transparency in digital media quality." Benchmark. Trust. Transparency. Three words, and IAS has arranged its entire existence so you cannot independently check a single one of them.

And the scale is genuinely nuts, I'll give them that. By their own telling, they chew through 280 billion digital interactions a day. They classify something like fifty years of video every twenty-four hours. They are one of two companies — the other being DoubleVerify — that the entire industry has agreed to trust as the referee. Was the ad seen by a human. Was it somewhere safe. Was it in the right country. IAS is one of maybe two neutral parties the whole league pays to make the call.

Here's the problem with being the referee. A good referee can show you the replay. Gets challenged, walks to the monitor, points at the freeze-frame, says here, look, this is why. That's what accountability looks like when it's real. It's checkable. It survives a skeptic staring right at it. It survives, say, an Orthodox Jew with receipts.

IAS mostly cannot do this. And where it could, it won't. What you get instead is a dashboard, some aggregate KPIs, and a very confident 99.99%. What you never get is the math. How a site gets scored. Where the thresholds sit. How often it's wrong, broken out by category. A single independent check of the live system against what actually happened. You are asked to trust the trust machine. On faith. The faith is the feature.

Real confidence publishes its error bars. Real confidence walks in the room and says here's where we fail, here's how often, here's what we're doing about it. That's how grown-up measurement talks, whether it's a drug trial or a fraud model. You lead with your limitations because your limitations are the most believable thing you've got.

IAS leads with four nines. And when somebody pushes back on the four nines, they don't open the machine. They post a blog.

Which brings us to the part where the industry stopped being polite about it. For a couple of years now, researchers and reporters have been doing the thing IAS never does — actually looking — and what they keep finding is major-brand ads sitting cheerfully next to exactly the content those brands pay IAS to keep them away from. Ads for companies you've heard of, next to content you would not want to describe to those companies' boards. On pages where IAS's own code was right there in the markup, supposedly on duty. This got the attention of United States senators. It got the attention of the Justice Department. And now it's got the attention of the buyers — which is the one that actually threatens the revenue.

And IAS's answer, every single time, has the same lovely shape. When the tool gets credit, it worked. When the tool failed — well, you configured it wrong. Or you misread the JavaScript. Or the page didn't hit the threshold. Or, my personal favorite, the tag you saw was for measuring, not blocking, so technically nobody promised you anything. Heads, IAS is your trust layer. Tails, you're holding it wrong. It's a coin welded to land on "not our problem," and that weld is the actual innovation.

Now here's the move that should make anyone still writing IAS a check put the pen down.

In September, IAS agreed to be taken private by a private-equity outfit called Novacap. All cash. About $1.9 billion. Ten bucks and change a share. Deal closed just before the end of the year. The stock came off the Nasdaq the next day. Vista Equity — which had owned the thing since 2018 and floated it in 2021 — took the money and left.

Sit with the timing. At the exact moment when independent, checkable, adversarial measurement has never mattered more — when AI is flooding the web with synthetic slop, when platforms like X have gotten toxic enough to run real accounts off of them — one of the two companies the industry trusts to measure all this just removed itself from the only venue on earth that forced it to disclose anything. No more annual filings. No more earnings calls where some analyst gets to ask a rude question on the record. No more audited numbers for a reporter to lay next to the marketing and start circling in red. The transparency company went dark. Let that irony finish cooking.

I'll be fair, because ADOTAT isn't cheap: going private is not a crime, and IAS has a tidy little story about why they did it, which I'll get to. But it changes the physics of accountability, and it's a very particular thing to do right after a brutal one-day stock collapse and a securities-fraud suit. Which — yes — we're getting to.

Three parts.

This one says the quiet thing out loud: the whole expensive machine does nothing, the near-perfect scores are marketing wearing a lab coat, and IAS built itself so you can never prove otherwise. I'm telling you that from inside the 99.99%, where it is loud and it is not safe.

Part Two is the autopsy. How the thing actually breaks, why "blame the client" is doing the work of ten lawyers, and why the failures aren't flukes — they're the design.

Part Three follows the money. Vista out, the crash, the lawsuit, and the brand-new Slack-and-Bumble CEO brought in to run the narrative faster than the classifier. Plus what publishers can actually do instead of renting a nap.

And here's the honest kicker, the reason this is a story and not a tantrum. IAS isn't collapsing. Nobody's fleeing. The big advertisers are renewing. It works well enough to keep selling the story. It just doesn't do the thing. And it has now made very sure you'll never get to check.

Let me do the thing IAS won't. Show my work, including the parts that could bite me.

What I said: that the "99%-plus safe, 97%-plus suitable" story IAS and X are selling about X is impossible as a description of the actual platform, that the whole apparatus is a sealed box with no error bars, and that going private turned off the last light anyone could read by.

Here's what holds up, cleanly, in the public record. X and its measurement partners do publicly claim brand-safety scores above 99% and suitability over 97%, with IAS and DoubleVerify cited as the "independent validation." That's not me putting words in their mouths. That's the June 2025 Brand Safety Score rollout, in their own press. The Adalytics research is real and well-documented: major brands, Microsoft, Meta, Disney, landing next to porn and racist content on pages where IAS or DV code sat right there in the markup. IAS's defense, that the report "inaccurately represented" its tech and mixed up publisher and advertiser tags, is real too, which is convenient, because it's exactly the "you're holding it wrong" move I told you to watch for. DoubleVerify went further and called the findings "entirely manufactured," which is a fun thing to say about screenshots. Senators Blackburn and Blumenthal sent letters. The Novacap take-private ($1.9 billion, $10.30 a share, roughly a 22% premium, closed December 23, delisted the next day, Vista out) is all boring corporate fact. And the box is sealed: nobody publishes false-negative rates by category, thresholds, or a live-classifier audit against log-level reality. Not IAS, not X, not DV.

Now the part where I stay honest.

Plank one, the wobbly one. The pristine "99.99%," the four nines I keep swinging at, shows up in marketing, but the cleanly sourced public number is "above 99%," which is a different animal only in that it's vaguer.

Either way you can't reconstruct the denominator, so the criticism survives. But if you're a lawyer looking for a thread to pull, that's the thread.

Plank two, the one to nail down. The "41% single-day collapse plus securities-fraud class action" is real in the litigation and ticker record, but it lives in paywalled dockets, not the open trade press. It's a plank, not a slab. Source it hard before it runs.

And here's the honest ceiling on the whole thing: "the number is impossible" is an argument, not a lab result. Nobody has run the adversarial, log-level audit that would falsify IAS's rate under IAS's own secret definitions, because IAS won't hand over the definitions. That's not a gap in my reporting. That's the crime scene. The reason I can't prove the number wrong is the same reason you can't prove it right: the people selling "transparency" sealed the box.

So what's the actual "evidence against me"? I went looking for it. Genuinely.

Here's the full inventory. It is (a) platform and vendor self-reports, the marketing grading its own marketing, and (b) the sealed box itself, the fact that the number can't be disproven because it can't be inspected.

That's it.

That's the defense.

"You can't prove us wrong" is doing all the work, and "you can't prove us wrong" is not the flex they think it is. Even the academics land here: automated suitability tools solve a real scale problem, but they are structurally incapable of the near-perfect truth the marketing implies, and the more automated and standardized they get, the more they bake bias, blind spots, and unfalsifiability straight into the buy. (Policy Review.)

One more receipt for the "even the refs blow it" file: DoubleVerify admitted it misreported X's brand-safety rate on its own dashboard for nearly five months in 2023 and 2024, a "graphical error," while everyone stared at a wrong number and nobody caught it. If the four-nines crowd can be wrong for a season without anyone noticing, forgive me for not treating the next four nines as gospel.

And the tell that tells the most: I could not find a single expert willing to defend IAS on the record. Not one. Plenty agreed with me. Almost none would say so with their name attached, which in this industry is its own kind of answer. I even did the dumb, honest thing and ran the whole argument through ChatGPT, Claude, and Perplexity, which proves exactly nothing, they're agreement machines with good manners, but for what it's worth, not one of the three could find real evidence against it either. Same conclusion, four times, three of them made of math.

Bottom line, the fair version: nothing in the public record today cleanly refutes the core of this. The "evidence against me" is self-reports and silence. Which, if you've been paying attention, is the whole point.

How we reported this. ADOTAT tested a claim the verification industry makes constantly and proves rarely: that automated brand-safety tools deliver near-perfect protection. We built this three-part series on IAS and Novacap deal materials and the December delisting, the Vista exit, IAS's February 2024 earnings and guidance, the July 2026 CEO change, IAS's own marketing and "commitment to media quality" posts, the securities complaint and market data for the one-day drop, the Adalytics reports and their trade coverage, the Blackburn-Blumenthal letters, DoubleVerify's disclosed dashboard error, and the Policy Review work on the limits of automated suitability tools; the author's account of harassment on X is firsthand testimony, marked as witness, not data. We kept record, reporting, and our own argument clearly separate, and where the data doesn't exist publicly, because IAS won't release it, we say so rather than treat an absence as proof. We could not independently falsify IAS's X safety numbers, because IAS discloses nothing that permits replication, which is the center of the story, not a hole in it. IAS was contacted before publication and offered a full right of reply; CEO Lidiane Jones and advisor Lisa Utzschneider were invited to comment. ADOTAT has no financial relationship with IAS, DoubleVerify, Novacap, Vista, or Adalytics and holds no related position; the author is personally targeted with identity-based harassment on X and discloses that stake on purpose.