OpenAI Engineer: Users Won’t Click Links. Microsoft’s Copilot CTR Gap for B2B SaaS

An OpenAI engineer put the zero-click problem in one sentence: show the links as hard as you want, users still will not click.

That line, plus Microsoft click-through ranges that put Bing Chat CTR far below classic Bing search for major publisher sites, just moved from sealed discovery into the public record. Search Engine Journal amplified the publishers’ brief overnight. For B2B SaaS teams that celebrate ChatGPT mentions as “traffic,” the filing is a measurement wake-up call from the platforms’ own documents.

---

What the publishers’ brief claims

On September 17, 2026, The New York Times, Daily News titles, Ziff Davis, the Center for Investigative Reporting, and The Intercept filed a public combined summary judgment brief in the Southern District of New York. PPC Land’s deep read walks the 92-page record: training copies, grounding, paywall issues, and the traffic evidence.

The traffic section is what operators should screenshot for leadership.

“no matter how prominently we show the links, users won’t click.”

That is the OpenAI engineer quote the brief attributes to February 2023 internal messaging, as reported by Search Engine Journal via The Times’s coverage of the filing.

Turley’s quoted lines go further: once ChatGPT’s browse function answered, there was “no good reason to click” the original source, and OpenAI products are “largely substitutive, period.”

Microsoft marketing language in the brief points the same direction. Copilot’s home page, per the publishers, greeted users with talk-through-answers framing instead of click-through-links framing.

The CTR ranges (Microsoft data, as cited)

Publisher group	CTR in Bing Chat vs Bing Web Search
The New York Times sites	87% to 93% lower
Daily News plaintiffs	83% to 91% lower
Ziff Davis sites	51% to 94% lower

Read these as alleged comparative CTRs from Microsoft measurement inside litigation, not as your SaaS site’s forecast. The filing does not publish query counts, date ranges, or how the low/high ends of each band were chosen. Microsoft renamed Bing Chat to Copilot in late 2023; the product name alone does not date the sample.

OpenAI and Microsoft filed their own summary judgment motions on September 4 and argue fair use and non-substitution. Judge Sidney H. Stein set opposition and reply deadlines into November 2026. Nothing here is a final judgment.

---

What this measures that GSC still does not

Microsoft’s AI performance surfaces in Bing Webmaster Tools already separate citations from classic search metrics. Google’s Search Console generative AI report shows impressions for AI Overviews and AI Mode, not AI-only clicks.

The publishers’ brief fills a gap both dashboards leave open: once a link appears in an answer, how often does anyone click it versus how often the same URL earns a click in classic search?

That is the step after citation. DerivateX already framed the Google side of this gap when AI Overview links began testing into AI Mode. This filing adds OpenAI/Microsoft-side evidence for the same structural point: answer engines are built to resolve the question in-product.

---

Why B2B SaaS should care (even if you are not a newsroom)

News publishers are the plaintiffs. Your buyers still use the same products.

1. Mention ≠ session. A ChatGPT recommendation can shape a shortlist while GA4 shows flat “chatgpt.com / referral.” That is consistent with ChatGPT owning the lion’s share of measurable AI referrals while most answer journeys never create a session.
2. Citation KPIs without exit KPIs lie. Rising citation share with flat demos is a board story waiting to go wrong. Keep recommendation share vs citation share as separate lines.
3. The rare click must carry buying facts. When someone does leave Copilot or ChatGPT for your URL, the first screen needs pricing bands, limits, security/SSO truth, and proof the model summary skipped. Thin “we got cited” pages waste the exit.

---

Operator playbook this week

Layer 1: Split the scoreboard

1. Prompt-panel naming across ChatGPT, Copilot, Perplexity, Gemini, Claude (same 15–20 buyer prompts, three runs each).
2. Referral sessions tagged by AI host (and Bing/Copilot paths where you can see them).
3. CRM “how did you hear” with explicit AI options.

If Layer 1 rises and Layer 2 does not, you are winning the answer, not the visit. That can still be fine if Layer 3 moves. It is not fine if finance thinks Layer 1 is traffic.

Layer 2: Instrument the leave-AI click

• Compare demo starts from AI referrals vs organic for the same commercial URLs.
• On money pages, put constraints and comparison tables above the fold.
• Do not “optimize for citation chips” by stripping the facts buyers need after the click.

Layer 3: Treat platform design as a given

Internal quotes from 2023 and Copilot marketing copy are not surprises. They describe a product that answers first. Plan distribution and proof for a world where most AI answers end without a session, then harvest the minority that do click.

---

Worked example: mid-market payments SaaS

Priority prompts: “best payment orchestration for SaaS,” “Stripe Connect alternatives for marketplaces,” “embedded finance platform for B2B.”

Weak read of this filing: “ChatGPT is killing our traffic; pause content.”

Better read:

1. Confirm whether you are named in those prompts even when referrals are thin.
2. Map which third-party sources (docs, G2, partner blogs) appear next to you.
3. Rebuild the commercial URLs that earn the rare Copilot/ChatGPT click so the first scroll answers packaging and risk questions the model softens.
4. Report to the CMO with three lines: naming rate, AI referral sessions, CRM AI-sourced opportunities. Never one blended “AI traffic” number.

---

How DerivateX uses defendant-side CTR evidence

1. Treat answer-engine visibility as a shortlist surface first and a referral channel second.
2. Refuse dashboards that collapse citations, impressions, and sessions into one vanity metric.
3. Build owned pages and third-party facts so the model can name you accurately, then so the exit click can convert.
4. Keep Google, ChatGPT, and Copilot in one weekly ritual. The court fight is about news copyright; the product behavior hits every B2B category.

If you need a prompt-level baseline across ChatGPT, Perplexity, Gemini, and Claude, start with a free AI visibility audit. Engagement options are on pricing. For proof patterns, see the Gumlet and REsimpli case studies. For the Google-side twin of this story, read AI Overview links opening AI Mode and the GSC generative AI report guide.

---

FAQ

What is new in the last day?

Search Engine Journal and PPC Land amplified the publishers’ Sept. 17 brief, including the OpenAI “won’t click” quote and Microsoft Bing Chat vs Bing search CTR ranges.

Are the CTR numbers proven for every site?

No. They are Microsoft comparative CTRs for named publisher groups as cited by plaintiffs. Methodology details are incomplete in the public reporting. Do not paste “−90% traffic” onto a SaaS board slide.

Does this mean AI citations are worthless?

No. Citations and recommendations still shape shortlists when sessions never happen. They are a weak proxy for pipeline. Measure naming, referrals, and CRM self-report separately.

How is this different from the Google AI Mode click story?

The Google piece covers Overview→AI Mode UI behavior and the Wang et al. field experiment. This piece covers OpenAI/Microsoft internal statements and Microsoft first-party chat-vs-search CTR bands from the publishers’ brief. Same structural lesson, different evidence and platforms.

Should we block AI crawlers?

That is a legal and distribution decision, not a default GEO tactic. Blocking can cut training/grounding exposure and also remove the citation path you may still want. Decide per engine with counsel and growth, not from one lawsuit headline.

What should we change this week?

Split AI naming from AI sessions in reporting, harden money-page first screens for the rare exit click, and re-run commercial prompt panels including Copilot where your buyers use it.