Copilot is testing a judge, not a writer

A decision-scoring model is a model that reads an input and a fixed set of options and returns a probability for each option, instead of generating text. Microsoft released one on October 9, 2026, and the Bing and Copilot team says it is already experimenting with it. The model will not write your summary in Copilot. It may decide which candidates are worth summarizing and whether the answer passed a quality check.

#What Microsoft shipped

Microsoft's announcement describes Microsoft-Decision-1 as a post-trained version of Qwen3.5-9B built for single-pass scoring. You give it a question and a closed set of answers (yes or no, multiple choice, a rating, or a rubric) and it returns a calibrated score for each. It is available in Microsoft Foundry and through OpenRouter.

The use cases Microsoft lists are the plumbing of an AI answer engine:

• Routing: deciding whether an agent continues, retries or hands off, and which model should take a request.
• Classification: intent analysis, content filtering, safety screening.
• Verification: judging AI responses against quality criteria.
• Prioritization: in Microsoft's words, assessing "how well a search result matches a query" and assigning a relevance label.

The economics are the point. Microsoft prices input at $0.042 per million tokens with output free, and reports median latency about 35 times faster than GPT-6 Sol. It also reports that the Copilot team used the model for quality control of chat and agentic responses, with quality competitive with GPT5.6 Luna at 100 times the speed. These are vendor figures without published per-benchmark accuracy, so treat them as claims, not measurements.

#What Copilot actually confirmed

Not much, and that matters. Search Engine Roundtable reported on October 12, 2026 that Rangan Majumder posted that Microsoft is "experimenting with it in Copilot." When asked whether Bing uses it, Krishna Madhavan replied that the team is "always experimenting." There is no statement that it runs for all queries or all users, and no statement about which decision inside Copilot it handles.

So the honest reading is narrow. A fast scorer exists, Microsoft built it for relevance labels and response checks, and Copilot is testing it somewhere. Nothing published says it selects citations today.

#Why a cheap judge changes the pipeline

Microsoft has used model-generated relevance judgments for years. In a SIGIR 2024 paper, Thomas, Spielman, Craswell and Mitra describe deploying LLM relevance labelling at scale at Bing, reporting labels about as accurate as human judges at a fraction of the cost, and better rankers trained on them. That work used a large general model offline to produce training labels.

A small scorer moves the same kind of judgment closer to the live request. When each check costs a fraction of a cent and adds little latency, an engine can afford to run more of them per answer: score more candidate passages, check more drafts before showing them, route more queries to cheaper paths. Microsoft's own illustration makes the latency case: 100 ms added to each of 20 sequential decisions adds two seconds to a workflow.

For a publisher, this means your page may be read several times by narrow judges before a generative model ever quotes it. Each judge asks a closed question. Does this passage match this query? Does this answer meet the rubric? Which of these candidates is most relevant?

#What a scorer rewards

We do not know the prompts or rubrics Copilot uses, so the useful guidance is the kind that holds for any closed-question judge.

A clear match beats a near match. A relevance scorer returns a label for a query and a candidate. A passage that names the entity, states the claim and gives the specifics in one place is easier to label as a match than a section that spreads the answer across three headings.

Wording tricks have less room. Microsoft reports that across eight input perturbations, decisions flipped on 1.3% of cases on average, with no flips when option descriptions were paraphrased or options reordered. If that holds in production, rephrasing a page to sound more relevant is unlikely to move a scorer. Adding the missing fact is more likely to.

Consistency gets checked. A verification model judging a response against quality criteria is looking for errors and unsupported claims. A page whose price, date or specification disagrees with its own structured data, or with the same brand's other pages, gives the checker a reason to prefer another source.

The retrieval phrase is the real query. Copilot does not score your page against what the user typed. It retrieves with its own shorter phrases. Bing Webmaster Tools exposes these as grounding queries, which Microsoft describes as the key phrases the AI used when retrieving content that was referenced in answers. Those phrases are the closest public proxy for the question a scorer asks about your page.

#What to do this week

1. Export grounding queries from Bing Webmaster Tools. Read them as the questions a relevance judge is asking. For each high-value phrase, check whether one passage on the cited page answers it directly. Remember that Microsoft calls the data "a sample of overall citation activity," so read counts as floors.
2. Fix the passages, not the phrasing. Where the answer is split or vague, consolidate it into one self-contained block with the entity name and specifics. Do not chase synonyms.
3. Diff facts across your own surfaces. Compare page copy, structured data, feeds and third-party profiles for the same product or claim. A verification step can penalize any disagreement it sees.
4. Annotate October 9, 2026 in your Copilot reporting. If Copilot citation counts or cited pages shift in the following weeks, a scoring change in the pipeline is one candidate explanation. It is not the only one, and there is no announcement that would confirm it.
5. Keep Copilot separate from other engines. A Microsoft scoring change says nothing about ChatGPT, Gemini or Google AI Overviews. Mixing them hides the signal.

#The takeaway

Generative engines are assembled from many decisions, and most of them are not generation. Microsoft has now released a model built only for those decisions and says Copilot is trying it. That does not tell you what Copilot cites tomorrow. It does tell you where the pressure is going: toward fast, repeated, closed-question checks of whether your page matches and whether it holds up. Pages that answer one question clearly and state their facts consistently are easier for any judge to pass.