Bing AI Training and Grounding Controls: What to Enable
If you run an ecommerce store or a brand site, turn AI Grounding on and leave it on. Grounding decides whether Bing's AI answers can retrieve your pages and cite them, and losing those citations costs you more than anything you'd protect by switching it off. AI Training is a separate decision. For a product catalog, we'd allow it. For a site whose original content is the product, we'd opt out. Limited Grounding is a narrow tool for proprietary or gated content. It shouldn't be anyone's default. None of these toggles should push you to block Bingbot, because that removes you from Bing search entirely.
The rest of this piece explains the reasoning and gives you a decision framework by site type. It closes with a list of what to audit before you touch the settings.
What the three controls actually separate
The Bing Webmaster AI training and grounding controls split one old question into three. The old question was whether Bing can crawl and index your site. The new questions are about what Microsoft's AI systems can do with what Bingbot already crawled.
- AI Training covers whether your content can go into the data used to build and improve AI models. Training happens once, upstream. Your page becomes part of what the model "knows," with no link back to you and no visit.
- AI Grounding covers whether Bing's AI experiences can pull your pages at answer time. When a user asks a question, the system retrieves current pages, generates a response from them, and cites the sources. This is how a product page ends up linked inside an AI answer.
- Limited Grounding is the middle setting. Going by Microsoft's framing, it permits retrieval with restrictions on how much of your content the answer reproduces.
Here's where we could be wrong. Microsoft's exact definitions, and what "limited" permits in practice, are set by its own documentation, and that documentation can change. Read the current Bing Webmaster Tools help text before you change anything. If Microsoft defines Limited Grounding as something that still cites and links in full, our advice on it shifts. Our recommendation on Grounding and Training stays the same.
The distinction that matters is timing and attribution. Training is upstream and anonymous. Grounding happens in real time and comes with a citation. For a site that sells things, those two have very different economics.
Why Grounding is the setting you protect
Grounding is retrieval, and retrieval is where search ranking still does the work. An AI answer that's grounded in live web results draws from the pages the search engine already considers relevant. If you opt out, you're removing your pages from the candidate pool at exactly the moment a buyer asks a question your catalog answers.
We have direct evidence of how tightly retrieval-based citations track rank, though it comes from Google rather than Bing. Between July and September 2026, we analyzed 2,350 tracked keyword snapshots across 314 keywords on client sites we manage, all on results pages where a Google AI Overview appeared. The AI Overview cited the site:
| Site's organic position | Share of snapshots where the AI Overview cited the site |
|---|---|
| Position 1 | 73% |
| Positions 2–3 | 49% |
| Positions 4–10 | 42% |
| Positions 11–20 | 9% |
| Beyond 20 or not ranking | 3% |
The pattern held on nearly 80% of the sites. The rest didn't have enough data to judge. This is a correlation, not proven causation, and Bing's system is a different system. The mechanism carries over, though. Grounded answers pull from pages the engine already trusts for the query. We walk through the full dataset in AI Overview Citations Follow Rank: What Our Data Shows.
Two details from that analysis bear on the Bing decision.
About a quarter of citations went to a different page on the same site than the page that ranked. The AI system picked a sibling page. Maybe it was a buying guide instead of the category page, or a spec-heavy product page instead of the brand landing page. That means you can't decide grounding page by page based on what ranks. Content you'd write off as "not our money page" can be what gets cited.
Citations without a top-20 rank were mostly comparison queries. "X vs Y" and "best X for Z" questions pull from a wider net. Ecommerce and brand sites compete hardest on exactly those queries, and comparison content is often what a store has invested in most. Opting out of grounding hands those answers to whoever stayed in.
So the case for Grounding is straightforward. If you rank on Bing, grounding is how that ranking turns into visibility inside Bing's AI answers. Switching it off doesn't protect your content. It only takes away the citation.
Training: the trade-off you should actually weigh
Training is the harder call, because the cost is real and the benefit is indirect.
Here's the cost. Content that goes into a model's training data can be paraphrased back to users with no citation and no click. The model absorbs what you wrote and gives you nothing visible in return.
The benefit is that models trained on your content "know" your brand, your product names, and the facts about your category. When someone asks an assistant about a product type you're known for, a model that's seen your pages has a better chance of describing you accurately.
The deciding question is whether the value of your content comes from it being exclusive or from it being widespread.
For a product catalog, widespread wins. Your specs, dimensions, materials, compatibility notes, and use cases already exist so that buyers can find them. You publish them in product feeds, on marketplaces, and in structured data. Holding them back from training protects nothing a competitor couldn't read off your product page. Allow training.
For original content that you monetize, exclusive wins. Proprietary research, paid guides, deep editorial work that earns links, a methodology you sell: if a model can reproduce that content's value without sending anyone to you, training works against you. Opt out.
For a brand site with mixed content, the catalog logic usually wins. Brand pages, about pages, and product education exist to be repeated. If you have one or two sections of proprietary editorial work, that's an argument for treating those sections differently, if the controls allow it, rather than opting the whole domain out.
One more point keeps people from overthinking this. Opting out of training on Bing doesn't remove your content from training data that's already been collected, and it doesn't govern other companies' crawlers. It's one control on one company's pipeline. Make the decision on its own merits rather than treating it as a blanket defense against AI.
Where Limited Grounding fits
Limited Grounding exists for content that should show up in answers without being substituted by them. Think of a recipe site, a gated research library, or a subscription knowledge base. You want the AI answer to point at you, and you don't want it to reproduce enough that the user never clicks.
For a typical ecommerce site, that problem doesn't exist. A grounded answer that summarizes your product page still ends with the buyer needing to come to your store to buy. The answer can't fulfill the order. Restricting how much of your product page gets used makes your listing a weaker candidate next to competitors who allowed full grounding, and you get nothing back for it.
Our position: use Limited Grounding only on content where the full answer would replace the visit. For catalogs, category pages, and brand pages, use full Grounding.
A decision framework by site type
| Site type | AI Training | AI Grounding | Limited Grounding |
|---|---|---|---|
| Ecommerce catalog (products, categories, brand pages) | Allow | Enable | Not needed |
| Ecommerce with a content hub of buying guides and comparisons | Allow | Enable | Not needed: comparison content is your citation engine |
| Brand site, product-led | Allow | Enable | Not needed |
| Brand or B2B site with proprietary research you sell | Opt out | Enable | Consider for the proprietary sections |
| Subscription or gated knowledge content | Opt out | Limited | Yes |
| Lead-gen service site (local or national) | Allow | Enable | Not needed |
The pattern is simple. Grounding stays on everywhere except where answers would fully replace the visit. Training stays on wherever your content's value comes from distribution rather than exclusivity.
What not to do: don't block Bingbot to get out of AI
The most expensive mistake here is reaching for robots. txt. Bing crawls with Bingbot, and blocking it to keep content away from AI also pulls those pages out of Bing's search index. You'd give up rankings, Shopping visibility, and every click Bing sends, all to make a decision the Webmaster Tools controls already let you make without that collateral damage.
The same applies to robots meta directives. Blanket noindex or restrictive snippet directives added "to stop AI" will cost you search visibility on every engine that honors them, Google included. If you've added any of these in the past year as an AI defense, review them against what the new controls now handle.
Audit before you toggle
Changing a setting in Webmaster Tools is the easy part. Making sure your site actually reflects what you decided takes more work. Here's what we check.
1. What robots directives your pages actually emit. Check the rendered page, not the template or the CMS config screen. On Magento stores, robots meta tags can come from the core design configuration, from category- and product-level overrides, and from extensions. We regularly find 40 to 70 extensions on stores that need maybe 15, and any of them can inject or override head tags. Crawl the site and read what's in the HTML Bingbot receives.
2. Whether Bing can reach the pages you want grounded. Grounding can only retrieve what Bing has crawled and indexed. In one technical SEO audit, we found crawl issues on 12,000 orphaned product URLs, which were product pages with no internal links pointing at them. Pages like that are invisible to retrieval no matter what you set in Webmaster Tools. Check Bing's index coverage against your actual catalog. Our technical SEO audit guide covers what a proper crawl review includes.
3. Canonicals and duplicate product URLs. If a product exists at four URLs because of layered navigation parameters, the grounded answer may cite the weakest one, or none. Clean canonicals give retrieval one clear page to pick.
4. Structured data on product pages. Price, availability, brand, and identifiers in Product schema give an AI answer facts it can state with confidence. Missing or stale schema makes your page a worse source than a competitor's.
5. Platform floor. If any part of your store still runs on Magento 1, AI settings aren't your priority. Magento 1 reached end of life in June 2020 and gets no security patches. We still find stores running Magento 1, completely unpatched, and we treat any of them as already compromised. Moving to Magento 2 comes first.
6. Your content sections, mapped to the framework above. List every section of the site: products, categories, guides, comparisons, research, gated content. Assign each one a training and grounding decision. Most ecommerce sites end up with one answer for the whole domain. Content-heavy brands sometimes end up with two.
Measure what changes
Once your settings are in place, track the result rather than assuming it.
Watch Bing organic clicks and impressions in Webmaster Tools for the sections you care about, so you can confirm your search visibility held steady. Watch referral traffic from Copilot and other Microsoft AI surfaces in your analytics. Expect some of it to be undercounted. AI assistant traffic often arrives without a clean referrer and shows up as direct, a problem we break down in ChatGPT Traffic Showing as Direct in GA4. The same fixes apply to Microsoft's surfaces. Then spot-check the answers yourself. Run the comparison and "best for" queries in your category and see who gets cited.
If citations drop after you change a setting, your configuration is the first suspect. If citations never show up even with grounding on, the cause is usually rank or crawlability. That brings you back to the audit list above.
How we handle this for the stores we run
When we run SEO and development for an ecommerce site, these settings go through the same review as every other change that touches crawling. We define the policy per content section, check that the rendered pages match it, confirm that Bing's index covers the catalog, and track citations and clicks afterward. Because strategy, SEO, and Magento development sit with one team, the person who decides the AI policy is working alongside the people who can fix the extension emitting the wrong robots tag. Nobody has to hand off a ticket and hope it gets done.
AI visibility is also one piece of a bigger operational question. If you're working out how AI should fit into your catalog operations, your content pipeline, or your order workflows, that's what our AI implementation work for commerce and operations covers.
The short version
- AI Grounding: enable it. Grounding is how your Bing rankings turn into citations inside AI answers. Retrieval-based citations follow rank, and comparison content, which is where ecommerce competes hardest, depends on it.
- AI Training: allow it for catalogs and brand content, opt out for content you sell. Base the decision on whether your content's value comes from exclusivity or from distribution.
- Limited Grounding: only for content where the answer would replace the visit. Ecommerce product pages rarely fall into that category.
- Never block Bingbot to get out of AI. You'll lose search visibility, and the Webmaster Tools controls make that sacrifice unnecessary.
- Audit the rendered site before you trust a setting. Extensions, orphaned URLs, and duplicate product pages decide what Bing can actually retrieve.
If you want a second set of eyes on your Bing setup, from robots directives to index coverage to the citations you're actually earning, get an audit. We'll show you where your catalog stands in Bing search and in AI answers, and what to fix first.