Microsoft-Decision-1 in Bing and Copilot: What AEO Teams Should Do

Microsoft-Decision-1 is a decision-scoring model that Microsoft says is already being tested in Bing Search and Copilot, as Search Engine Roundtable reported. For AEO and GEO teams the near-term change is small, but the direction matters: answer engines are adding fast layers that weigh candidates against each other before anything is written or cited. The useful response is to make your pages easier to compare and verify, and to stop treating each engine's experiment as a new ranking factor.

Key takeaways

• Search Engine Roundtable reports that Microsoft announced Microsoft-Decision-1 last week and that it is already being experimented with in Bing Search and Copilot.
• A decision-scoring model works on choices between options, which puts it between retrieval and the finished answer. That is where citations are won or lost.
• Scorers favour facts that are consistent and comparable. Persuasive copy matters less at this stage than a clean, matching set of attributes.
• Start with one fact sheet per product, plan or entity, then make every page, profile and partner listing agree with it.
• Measure with a fixed prompt panel, a fact-consistency audit and consistently tagged Bing and Copilot referrals. Results will be noisy, so track share over many prompts.
• Avoid rebuilding your site around one model name. Experiments change, and the fundamentals do not.

What Microsoft announced, and what it leaves open

Search Engine Roundtable reports that Microsoft announced Microsoft-Decision-1 last week and that the model is already being experimented with in both Bing Search and Copilot. Microsoft describes it as a decision-scoring model with top performance in latency and quality on structured decision tasks, and says it outperforms both large language models and other decision models.

Read that with care. It is Microsoft's claim about its own model, on tasks Microsoft chose. Microsoft's own announcement is the place to check which tasks those are and how they were measured. 'Experimenting' also has a narrow meaning: some queries, some users, some surfaces, for a period nobody outside the company knows. You may never see it in your own results, and it can change or vanish without notice.

What you can say with confidence is the shape of the thing. This is not another chat model that writes prose. It is a model built to score options for a decision, quickly and cheaply enough to run inside a search request. That is a different job from the one language models do, and it is the stage of the pipeline that AEO teams have thought about least.

Why a decision-scoring model matters for AEO and GEO

Most GEO advice concentrates on the two ends of the pipeline: get retrieved, get quoted. A scoring model sits in the middle. Somewhere between a pool of candidate pages, products or passages and a finished answer, something has to pick. Which three vendors make the list? Which of two conflicting pages is believed on price? Which passage gets quoted?

Using a large language model for every one of those choices is slow and expensive. A compact model trained to score structured choices is the obvious fit, and that is the case Microsoft is making with its latency and quality claim. In what has been reported, Microsoft has not said which decisions Bing and Copilot are handing to it. Treat everything that follows as inference about direction, not a description of a mechanism.

The inference is still useful. Scoring works best on things that can be compared: attributes with consistent names, values that agree across sources, claims that attach cleanly to a named entity.

Take a company selling workforce scheduling software. Its pricing page says 'from $6 per user', a partner listing says '$9', and a review site says 'custom pricing'. Any scorer now has three conflicting inputs and no reason to prefer yours. Compare that with a competitor whose plan names, limits and prices are identical everywhere. It has done nothing clever. It is simply easy to score.

That is the practical shift. At this stage of the pipeline, being legible as a set of facts counts for more than being persuasive.

How AI answer engines select and cite sources

Strip away the vendor differences and most answer engines follow a similar sequence. The details vary by product and change often, so check each vendor's documentation for what it states about its own system.

1. The prompt is interpreted and often expanded into several narrower searches. Ahrefs published a study of how ChatGPT does query fan-out that is worth reading for the mechanics. The point for you: one question becomes many retrievals, and your page may be matched to a sub-question you never targeted.
2. Candidate pages or passages are retrieved from an index or from live search.
3. Candidates are filtered and ranked on relevance to the sub-question, freshness, source trust and whether the passage actually answers it.
4. A model writes the answer from the passages that survive.
5. Some of those passages are attached as citations.

A decision-scoring model would plausibly live in steps three and five. If so, it changes the odds of surviving the filter, not the odds of being retrieved. A page that ranks but doesn't make its passages easy to lift gets retrieved and then dropped. That is the familiar case of a page that ranks well and never gets cited, and a scoring layer would sharpen it.

Two consequences follow. First, passage quality becomes a first-class concern: each section should answer one question completely, with the entity named, so it survives being quoted without its surroundings. Second, agreement between sources becomes a signal. If your claim matches what other credible pages say, it is cheap to accept. If it conflicts, it costs the system something to resolve.

Why Bing-side changes reach beyond Bing

Bing is easy to dismiss when your analytics show a small share of organic traffic. That undersells it. Bing's index has historically fed other products' search features, and ChatGPT's search has been reported to draw on Bing results among other sources. Each vendor's documentation states what it uses today, and that can change.

If Bing and Copilot get better at choosing between candidates, the effects can show up in places that never mention Microsoft. So a Bing referral number is the wrong yardstick. The question to ask is whether your key pages are crawled, indexed and clean in Bing's index at all.

That is cheap to check and cheap to fix. Verify your sitemap and robots rules, confirm the pages that carry your facts render without depending on client-side scripts, and use Bing Webmaster Tools. If your CMS supports IndexNow, which Bing supports, turn it on so updates reach the index quickly.

What to change in practice

None of this needs a new tool or a new team. It needs a few unglamorous habits applied consistently.

1. Write one fact sheet per entity. For each product, plan, location or person: official name, price, limits, availability, dates, who it is for. One owner, one place. Product marketing usually owns this best.
2. Make every surface match it. Site pages, app-store listings, partner directories, press boilerplate, social bios. Mismatches are the cheapest thing for a scorer to penalise and the cheapest thing for you to fix.
3. Lead sections with a standalone answer. Phrase the heading as the question a buyer would ask, then give a direct answer of a few sentences that makes sense when lifted out of the page.
4. Compare on the page, in a table. Use the same attributes for every option, including the ones where a competitor wins. An evasive comparison gives a scorer nothing to work with and a reader nothing to trust.
5. Mark up only what the page says. Structured data should mirror visible text. Schema for a claim the page does not make is worse than none.
6. Date and own your content. Show an updated date only when something changed, and name a real author or reviewer. Both help a human decide whether to trust the page, and they are cheap signals for a system too.

The trade-off most teams will hit is the honest comparison table. Sales will not want a competitor's strengths on your site. The counterargument is simple. A buyer asking an assistant for a shortlist is going to see those strengths anyway, and the page that states them accurately is the one most likely to be treated as a reliable source.

How to measure the effect

Measurement here is noisy. Answers vary from run to run, prompts are phrased differently by every user, and engines change without announcement. So measure share across a panel, not single wins, and repeat runs before drawing conclusions.

Build a panel of 30 to 50 prompts that mirror real buying questions, run them on a schedule across Copilot, Bing and the other assistants your buyers use, and log which domains are cited and how your brand is described. A platform like MediaPilot is built around AI-search visibility; the product page shows what it covers.

On the analytics side, referral traffic only helps if sources are named consistently. A tracking specification is the document that settles how Bing, Copilot and other assistant referrals are labelled, so that month-over-month comparisons mean something. Teams that own that layer will find the analytics teams page a reasonable starting point.

What to track	How	Owner	Cadence
Citation share on a fixed prompt panel	Run the same 30-50 buying prompts across engines, log cited domains and brand descriptions	SEO lead	Weekly
Fact consistency	Audit price, plan and spec values across site, profiles and partner listings against the fact sheet	Product marketing	Monthly
Bing and Copilot referral traffic	Segment by a consistent source and medium naming scheme in analytics	Analytics lead	Weekly
Indexation of fact-bearing pages in Bing	Review coverage and crawl reports in Bing Webmaster Tools	SEO lead	Monthly
Quality of AI-referred sessions	Compare engagement and conversion with organic search on the same landing pages	Growth lead	Monthly

If the fact-consistency audit is the only item you do this quarter, you will still have done the most useful one. It also needs no model access and no new budget.

What to ignore and what not to do

Experiments like this attract speculation quickly. A few things are worth refusing.

• Don't build a strategy around a model name. Microsoft-Decision-1 may be replaced, renamed or confined to a few query types. The habits above survive any of those.
• Don't claim you can optimise for a scoring model. Nobody outside Microsoft knows what it weighs. Anyone selling a Decision-1 playbook is guessing.
• Don't publish a wave of thin comparison pages. Fifty near-identical 'X vs Y' pages dilute your facts and give a scorer more conflicting copies of the same claim.
• Don't hide pricing or limits to protect negotiation. If the facts are missing, a system will use someone else's version of them.
• Don't stuff structured data. Markup that goes beyond the visible page invites distrust and does not make a weak page citable.
• Don't read one run as a trend. One prompt on one day proves nothing. Look at share over a panel, over weeks.

The broader point is that AEO is increasingly an operations problem as much as a content one. Keeping facts consistent across a site, a CMS, feeds and a dozen third-party profiles is coordination work, and teams that treat it as such are the ones that get chosen. The what is marketing operations guide covers that discipline if you are building the case internally.

Frequently asked questions

What is Microsoft-Decision-1?

It is a decision-scoring model from Microsoft. According to Search Engine Roundtable, Microsoft says it delivers top latency and quality on structured decision tasks and outperforms both large language models and other decision models. The same report says it is already being experimented with in Bing Search and Copilot. For the task definitions and benchmarks, go to Microsoft's announcement.

Do I need a separate Bing and Copilot strategy?

Not a separate content strategy. The same fact-consistent, answer-first pages serve every engine. What you do need is confirmation that Bing can crawl and index your key pages, because Bing's index and Microsoft's assistants are where this experiment is happening.

Will structured data help me get chosen?

It helps when it mirrors what the page visibly says, because it makes entities and attributes unambiguous. It does not rescue a page that conflicts with other sources or buries its answer. Treat it as a clean-up step after the fact sheet and the on-page answers, not a substitute.

How soon will I see any effect?

Possibly not at all, and not in a way you can isolate. The tests are limited and the engines change often. You can expect fact fixes and indexing improvements to pay off on a timescale of weeks as pages are recrawled, but attribute any gain to the whole set of changes rather than to one model.

Sources

• https://www.seroundtable.com/bing-microsoft-decision-1-42270.html
• https://ahrefs.com/blog/query-fan-out-chatgpt/
• https://www.conductor.com/academy/aeo-geo-benchmarks-report/