"No matter how prominently we show the links, users won't click." An OpenAI software engineer wrote that in February 2023, and it has now surfaced in a court filing.
Most marketers spent the last two years hoping that sentence was wrong. The people building the product apparently knew it was right from the start.
Matt Southern at Search Engine Journal walked through the brief this week. It comes from the news publishers suing OpenAI and Microsoft, filed on September 17 in the consolidated copyright case in New York. It is one side's argument and the court has not ruled.
The quotes are not the interesting part. The numbers are.
What the Microsoft data actually shows
The brief cites Microsoft data comparing click-through rates for the same publisher sites in Bing Chat and in Bing Web Search. For The New York Times, CTR was 87 to 93 percent lower in chat. For the Daily News group it was 83 to 91 percent lower. For Ziff Davis sites the range ran from 51 to 94 percent lower.
The same filing quotes Nick Turley, head of ChatGPT, saying there was "no good reason to click" once the answer had been given. That is not a bug report. That is a product philosophy, stated plainly by the person responsible for the product.
The caveats are real and worth keeping. The passage does not say when the data was collected, how many queries it covered, or what the low and high ends of each range represent. Bing Chat became Copilot in November 2023, and nothing ties these figures to the current product.
Even so, the direction is not in question. Across three very different publisher groups, the floor of the decline is 51 percent. There is no version of this range in which chat sends you the traffic search used to.
Cited and clicked are different measurements
This is where most reporting I see goes wrong. Teams track "AI visibility" and quietly assume it behaves like search visibility, with traffic following a few steps behind.
It does not. Microsoft's own AI Performance report in Bing Webmaster Tools counts citations and says plainly that those metrics differ from rankings, clicks and traffic. When it added Citation Share in June, it said the metric does not represent traffic share. Google's generative AI report in Search Console shows impressions and no clicks at all.
I wrote about that gap when Search Console started reporting AI impressions without clicks. This filing is the other half of the picture. The platforms are not hiding the clicks. For the most part, the clicks do not exist.
So you now have two separate outcomes. One is being named in the answer. The other is someone arriving at your site. They correlate weakly, and the correlation is getting weaker every quarter.
Treating them as one number is how a marketing team reports rising AI visibility in the same quarter that inbound leads fall. Both statements are true. Neither explains the other.
The answer is the shelf now
If the click is not coming, the answer itself has to do the selling. That changes what a good page is for.
For fifteen years, the job of a page was to earn the click and then convert the visit. In a chat interface, the page is raw material. It gets read by a machine, compressed into two sentences, and presented without your design, your proof points or your call to action.
What survives that compression is the claim, the number and the name. If your differentiator only makes sense after three scrolls and a demo video, it does not survive. If it fits in one sentence with a figure attached, it has a chance.
That is why I keep telling clients to rewrite their core pages as if a sceptical analyst will summarise them for a buyer who never visits. Because that is now the literal operating reality.
It also changes how you value a citation. A mention with your brand name, a specific capability and a reason to pick you is worth something even with zero clicks. A mention buried in a list of eight vendors, with no differentiator attached, is worth close to nothing, whatever the dashboard says.
Budget for the two outcomes separately
Here is the operational change. Stop reporting AI visibility as a traffic channel. Report it as a brand channel, closer to PR than to SEO.
Measure where you appear, what the answer says about you, and whether that description matches what you sell. Then measure the downstream effects that do not need a click: branded search volume, direct traffic, and the share of sales calls where the buyer already knows your name.
That is the job GEOflux.ai was built for. It maps where a brand shows up in AI-generated answers and why, which is the only part of this you can actually manage when the click is optional.
Keep your traffic plan, but fund it from the channels that still send traffic. Paid search, email, partnerships and communities you own did not stop working because chat arrived. They just stopped being subsidised by organic search.
And plan for the judge's timeline. Oppositions are due October 23, replies November 20. Whatever OpenAI and Microsoft say about these numbers will be the most honest public description of chat-driven traffic we are likely to get this year.
The engineer was right in 2023. The only question left is how long it takes your reporting to catch up.