Key takeaways
- Copilot and Meta AI both pull from web indexes, so a manual audit starts with prompts, but the fixes start with Bing indexing and Meta crawler access.
- Run about 12 prompts across three types (unbranded, branded, competitor comparison) and record presence, position, competitors named, cited URLs, and tone.
- Copilot's citation counts swing wildly. Repeat the same prompts on different days before drawing conclusions.
- Meta AI answers in private chats can't be observed, so your audit is a sample, not a census.
- Fifteen minutes gets you a baseline. Tracking it over time needs a tool.
Why these two assistants deserve their own audit
Most brand audits start and end with ChatGPT. That's understandable, but Microsoft Copilot sits inside Windows, Edge and Microsoft 365, and Meta AI sits inside WhatsApp, Instagram, Facebook and Messenger. Your buyers may be asking them questions without ever opening a dedicated AI app.
The two are also hard to measure. Copilot is volatile. Meta AI is mostly private. Neither gives you a tidy dashboard by default. So a manual audit is the honest first step, as long as you know what it can and can't tell you.
Promptwatch Data tracks citation behaviour across engines, and its Average Sources Per Response report shows why Copilot needs patience: its average has swung from under 2 sources per response to nearly 17 within a few weeks, then settled at the low end. ChatGPT sits around 5 sources per web-search answer, while Google AI Overviews and Perplexity sit near 10. Promptwatch reads the Copilot swings as Microsoft still reworking how it retrieves and attributes sources.
That has a direct consequence for you. A single run of a prompt tells you almost nothing.
Before you start: set up a clean test
Fifteen minutes is tight, so prepare three things first.
- A spreadsheet with columns for prompt, platform, date, brand mentioned (yes/no), position, competitors named, cited URLs, and sentiment.
- A logged-out or clean browser profile for Copilot, so your own history and Microsoft 365 context don't color the answer. If your buyers use Copilot at work, run a second pass signed in to a work account.
- A phone with Meta AI available, or the Meta AI web experience, depending on what is available in your country.
Note the country and language you test from. Answers differ by region, and your baseline is only useful if you can repeat it under the same conditions. The MindStudio guide to checking brand visibility makes the same point: don't run tests in your head, write them down, because the record becomes your benchmark.
Minutes 0 to 3: write your prompt set
Twelve prompts is enough. Split them into three groups of four.
| Prompt type | Example pattern | What it tells you |
|---|---|---|
| Unbranded commercial | "best [category] for [use case]" | Whether you are in the consideration set at all |
| Branded | "is [your brand] good for [use case]" or "[your brand] pricing" | What the assistant believes about you |
| Competitor comparison | "[your brand] vs [competitor]" | How you are framed against rivals |
Search Atlas recommends auditing unbranded commercial queries first, and I agree. Branded prompts almost always return you. The unbranded ones are where you find out whether you exist to a stranger.
Write them the way a real person would type them, not the way an SEO would. Add one or two with a location or budget if you sell locally or to a specific segment.
Minutes 3 to 8: audit Microsoft Copilot
Run each prompt in Copilot and fill in the spreadsheet row immediately. Don't batch the note-taking for later; you'll forget which citation belonged to which answer.
What to record for each answer
- Whether your brand is named at all, in the text or only in the source list.
- Where it appears. First mention, middle of a list, or an afterthought.
- Every competitor named.
- Each cited URL, and whether any of them are yours.
- Whether the description of you is accurate, outdated, or flattering.
Copilot usually answers with a short synthesis plus a handful of footnote citations. Because the citation count is unstable, treat a missing citation as a hint rather than a verdict.
Separate ads and shopping from organic answers
For product prompts, Copilot may show ads or merchant data. Microsoft Advertising has said Offer Highlights, such as free shipping or in-store pickup, appear in ads on product detail pages in Copilot, Edge and Bing, for retail in English-speaking markets. Structured product catalog data also feeds how products surface. Log those results in a separate column. A paid placement or a merchant feed match is not the same as being cited organically.
Why Bing matters here
Microsoft says its grounding is built on Bing and powers Microsoft AI experiences plus partner AI platforms that rely on Bing. The practical upshot: if your key pages don't rank or aren't indexed in Bing, don't expect Copilot to find them. Promptwatch makes the same recommendation, to check that Bing indexes and ranks your key pages well.
Search Atlas also points to the IndexNow protocol as a way to push new and updated URLs to Bing quickly instead of waiting for sitemap polling. That's a fix, not part of the audit, but note it if your pages turn out to be stale in Bing.
Minutes 8 to 12: audit Meta AI
Repeat the same twelve prompts in Meta AI. Keep the same columns.
Expect differences. A third-party vendor blog (Bingly, which sells a competing tracker, so take it with salt) claims Meta AI uses a web search partnership for current questions and shows citations less consistently than Perplexity. I couldn't confirm the current search provider from a primary source, so I wouldn't build strategy on that claim. What you can observe directly is whether Meta AI names you, whether it links anything, and what it says.
Two cautions specific to Meta AI:
- Much of its usage happens in private chats you can't see. Your audit samples what the assistant says to you, not what it says to everyone.
- It may draw on Meta's own platform data. It's worth checking that your business details on Facebook and Instagram match your website, particularly if you're a local or consumer brand. That advice comes from the same vendor blog and is plausible rather than proven.
Why Meta crawler access is worth checking now
Meta is crawling harder than it used to. In Promptwatch's Meta-WebIndexer report, Meta-WebIndexer went from about 2.2% of tracked AI crawler requests in mid-July 2026 to 37.8% on August 9, 2026, making it the heaviest AI crawler in Promptwatch's logs. A claim that Meta is building its own web index circulated on August 6, but the report treats it as unconfirmed. Promptwatch's own reading is that the crawler exists to improve Meta AI search results and cite sources.
In robots.txt and your server logs, look for these user agents:
meta-webindexer
meta-externalagent
meta-externalfetcher
FacebookBot
If you block them, you may be keeping your content out of Meta's index. Some sites blocked them because of server load, which is a reasonable call, but make it a deliberate one.
Minutes 12 to 15: score and compare
Now tally. For each platform, count how many of the twelve prompts mention you, how many cite your own domain, and how often a given competitor shows up.
| Metric | How to calculate | What a weak result suggests |
|---|---|---|
| Mention rate | Prompts naming you / total prompts | Coverage problem |
| Unbranded mention rate | Unbranded prompts naming you / 4 | You're not in the consideration set |
| Self-citation rate | Prompts citing your URLs / prompts naming you | Content or crawlability problem |
| Competitor frequency | Times each rival is named / total prompts | Who is taking your slot |
| Sentiment | Positive, neutral, negative, outdated | Reputation or stale-data problem |
Reading these together locates the problem. Low mention rate on unbranded prompts means you're missing from the shortlist. Mentions without your own URLs cited usually point at content, or at pages the assistant can't reach. Wrong details suggest outdated third-party sources are winning.
Repeat before you trust it
Given Copilot's swings, do the Copilot half again on two other days. If your brand appears in one run out of three, you have a fragile presence, not a visible one. Be wary of reading too much into a single week; Promptwatch's advice is to judge Copilot on monthly trends. A sudden jump or drop is more likely a retrieval change on Microsoft's side than something you did.
Tools that remove the manual work
The manual audit is a snapshot. It won't catch drift, and it can't run hundreds of prompts.
On the Microsoft side, Citations in Microsoft Clarity is now generally available. It unifies signals from Bing Webmaster Tools and Clarity, covering bot crawling, indexing, citations in AI answers and AI-referred traffic, with metrics such as Queries Cited, Citation Rate, Page Citations and Share of Authority. Microsoft is upfront that it's a representative view meant for trend analysis, not an exact log of individual answers. It's free and first-party, so set it up regardless. Bing Webmaster Tools also introduced an AI Performance report in public preview in February 2026; I haven't verified its details, so check Microsoft's documentation.
For cross-engine tracking, Promptwatch monitors the real interfaces of Microsoft Copilot and Meta AI alongside ChatGPT, Google AI Overviews, Gemini and Perplexity, and layers on crawler logs and AI-referred traffic so you can see whether a visibility change turned into visits. Its Content Agent and Action Board also help turn the audit gaps into published fixes.
Other trackers worth a look if you're comparing:
If you'd rather browse a wider field, the GEO software directory at bestgeosoftware.com lists more options. For the free first-party side, Google's equivalent is Search Console, and Bing's is Webmaster Tools.
Common audit mistakes
- Testing once. Copilot's source counts move too much for a one-day read.
- Mixing signed-in and signed-out results in the same spreadsheet.
- Counting ads and merchant feed placements as organic citations.
- Only testing branded prompts, which flatters you.
- Treating Meta AI results as population data. They're a sample.
- Changing the prompt wording every time, which makes trends meaningless.
What to fix first
Pick the cheapest fixes that match what your audit exposed.
- If Copilot doesn't cite you: confirm your key pages are indexed and ranking in Bing, add IndexNow, and make the first sentence of each section state a plain fact.
- If Meta AI doesn't know you: check the four Meta user agents in robots.txt and your CDN or WAF rules, and align business details across your site and Meta profiles.
- If competitors keep appearing instead of you: look at which URLs the assistants cite for them, then build or refresh comparable pages.
- If the description of you is wrong: find the outdated third-party source being cited and update it, or publish a clearer first-party page.
AI crawler traffic is also shifting. In Promptwatch's AI crawler traffic report, OpenAI made up 79.8% of verified requests in the week of Aug 31 to Sep 6, 2026, down from 94.8% in early June. If a provider has a big share in that report and nearly none in your logs, you're likely blocked.
The bottom line
Twelve prompts, two assistants, one spreadsheet, fifteen minutes. You'll learn whether you're in the shortlist, who's replacing you, and whether the facts are right. What you won't get is certainty, because Copilot's citations are unstable and Meta AI's real usage is private. Repeat the audit monthly, and when the spreadsheet gets tedious, move it into a tracker. For more on citation patterns across engines, Promptwatch's data hub is a good place to start.