The noai and noimageai meta tags are two of the most shared — and most misunderstood — AI opt-outs on the web. They look like ordinary robots directives, they take thirty seconds to add, and they promise exactly what many site owners want: keep my words and images out of AI training. The honest answer is that they are a polite request that the biggest AI companies do not document as supported, and they should never be your only opt-out. But the story does not end there. The same idea — controlling AI use page by page instead of site-wide — now has real, documented tools behind it, including Bing's new per-page training and grounding controls and a deadline that makes the old tags expire. This guide separates what works from what merely feels like it works, and gives you the page-level setup that actually holds up in 2026.
What noai and noimageai are
The two tags are simple. Added to a page's head section, they look like this:
<meta name="robots" content="noai, noimageai">The meaning is plain English: noai asks AI systems not to use the page's content for training, and noimageai asks the same for the page's images. They were introduced by DeviantArt in 2022, when artists discovered their work was being swept into image-generation training sets. The idea spread quickly through art communities, website builders and SEO plugins, and it is now one of the most commonly recommended AI opt-outs in circulation.
Its popularity is understandable. It is easy, it is per-page, and it feels like the robots meta tags that already control search engines. The problem is that feeling is not evidence. A meta directive only works if the system reading the page has agreed to honour it, and the agreement is the whole question.
Do Google and Bing respect noai and noimageai?
Short answer: no, not as documented controls. Google publishes a list of the robots meta directives it supports — noindex, nofollow, nosnippet, max-snippet, max-image-preview, notranslate, and a handful of others. noai and noimageai are not on it. To control Gemini's use of your content, Google documents the Google-Extended token in robots.txt, not a meta tag.
Bing's position is the same. Its documented robots directives do not include noai, and Microsoft directs site owners to robots.txt controls for AI training instead. Some dataset tools and artist-focused platforms have said they respect the tags, and that is worth something if your concern is a specific dataset rather than the major labs — but as a general opt-out against the companies building the models people actually use, the tags are not a control. They are a statement of preference that the intended audience has not agreed to read.
That distinction matters for how you spend your time. Adding the tags costs nothing and may help with the smaller operators that do honour them. Relying on them as your whole AI policy would be a mistake, because the largest collectors of training data do not list them as inputs.
The tags that do have documented support
The good news is that page-level AI control is not a fantasy — it exists, it is documented, and in Bing's case it now comes with a calendar. Here are the controls that major operators actually list.
Bing's noarchive and nocache — and their expiry date
Since 2023, Bing has read the noarchive and nocache robots directives as AI signals: content carrying them was treated as opting out of AI training and of grounding in Microsoft's AI answers. That reading ends soon. Microsoft announced on 30 October 2026 that noarchive and nocache will stop counting as training or grounding preferences on 1 March 2027. If your site relies on those tags today, they stop protecting you on that date unless you replace them.
The transition has a timetable worth putting in your calendar. On 15 December 2026, a temporary grounding preference opens in Bing Webmaster Tools, but only for sites that are already using noarchive or nocache on that date. On 1 February 2027, Bing starts reading the new preferences as it recrawls pages. On 1 March 2027, the old tags stop counting. On 30 September 2027, the temporary preferences expire. If you use the old tags, act before the March deadline — and note the December catch: the temporary setting is only available to sites already using the old tags when it opens, so removing them early would close that door.
Bing's new per-page controls: ai-training and ai-grounding
Microsoft replaced the old tags with two separate preferences, and this is the part that makes page-level control real. Each preference is set per page, with a meta tag or an HTTP header:
<meta name="Bingbot" content="ai-training:disallow">
<meta name="Bingbot" content="ai-grounding:disallow">The two preferences do different jobs. ai-training:disallow says the page must not be used to train models. ai-grounding:disallow says the page must not be used to ground Microsoft's AI answers — meaning Copilot and Bing's AI features should not draw on it when answering. Grounding also has a middle value, ai-grounding:disallowextended, which allows grounding in Microsoft-operated experiences only. The same values can be sent as an HTTP header, X-Robots-Tag, which is the right choice when you want the rule to cover files that have no HTML head section, such as PDFs.
Two details matter. First, the page-level setting overrides the site-wide robots.txt setting, so you can allow training site-wide and disallow it on individual pages, or the reverse. Second, neither preference changes eligibility for traditional Bing search results. Microsoft is explicit about that, and it removes the fear that opting out of AI use costs you search traffic. Microsoft does warn that choosing Disallow for grounding will significantly reduce your content's visibility in AI-generated experiences — which is the point, if that is what you want.
Google's nosnippet — a page-level control for AI Overviews
Google's documented page-level lever is nosnippet. Google's robots meta tag documentation states that the directive applies to all forms of search results — web search, Google Images, Discover, AI Overviews and AI Mode — and that it will also prevent the content from being used as a direct input for AI Overviews and AI Mode. That is a narrower promise than a training opt-out: it governs how the content appears in Google's search features, not whether it trains Gemini. For training, Google's documented control remains the Google-Extended token in robots.txt. But if your concern is a specific high-value page being quoted verbatim inside an AI Overview, nosnippet on that page is a documented, working lever — and it is one of the few page-level controls Google actually lists.
The site-level answer: the shared training opt-out
Page-level tags are the precision tool. The site-level tool is the shared robots.txt token that Google, Microsoft, OpenAI and Anthropic agreed to support on 30 October 2026:
User-agent: __usage_ai-training
Disallow: /One entry states a do-not-train preference to all four companies at once, instead of a separate stanza per crawler. It is still an IETF draft, not a finished standard, and it covers training only — not AI search answers. But it is the direction the industry is moving, and it belongs in your robots.txt alongside the per-crawler rules. Our guide to the shared opt-out covers the rollout timing and the limits in full.
A practical page-level setup, in order
Here is how to combine these pieces without contradicting yourself. The principle is simple: state the preference site-wide in robots.txt, then use page-level tags only where a page deserves different treatment from the rest of the site.
Step 1 — Set the site-wide position in robots.txt. Add the shared __usage_ai-training group and the named training-crawler stanzas for GPTBot, ClaudeBot, Google-Extended and the rest. If you are building the file now, the robots.txt generator produces it with the AI presets included, and the robots.txt tester confirms the result. This is your default: everything is opted out unless a page says otherwise.
Step 2 — Use Bing's per-page tags for the exceptions. If most of the site is open to AI use but a few pages are not — a pricing page, a proprietary research post, a licensed image gallery — add the Bingbot meta tags to those pages only. Because the page setting overrides robots.txt, a page carrying ai-training:disallow stays out of training even if the site-wide rule allows it. Use the header form for PDFs and other non-HTML files.
Step 3 — Add nosnippet where verbatim quoting is the concern. For pages whose value is in the exact wording — a definition, a data table, a short answer you do not want lifted into an AI Overview — add nosnippet. Understand the trade: it also removes the snippet from ordinary search results, so use it on pages where the snippet is not doing traffic work for you.
Step 4 — Keep noai and noimageai as a courtesy layer, not a control. Add them if you like; they cost nothing, they are recognised by some smaller operators and dataset tools, and they state your preference in plain sight. Just do not count them in your audit. When you verify your opt-out, verify the robots.txt rules and the Bing tags — the layers with documented support.
Step 5 — Replace noarchive and nocache before 1 March 2027. If your site uses those tags as AI controls today, schedule the swap now: move training control to the shared robots.txt token and the Bingbot meta tags, and check Bing Webmaster Tools on 15 December 2026 for the temporary grounding preference. After March, the old tags are search directives again, not AI controls.
What about images, PDFs and downloads?
Page-level meta tags only work on HTML pages, which is a real gap for the files AI systems value most: images and documents. Three options cover it.
For images, noimageai is the tag most often recommended — and, like its sibling, it is not documented as supported by the major operators. The stronger protection for images is at the server level: block the image-training crawlers from your image directories in robots.txt, and use the header form of the opt-out for image files. Our guides to protecting images from AI training and protecting PDFs and downloads cover the directory-level rules in detail.
For PDFs and other documents, the X-Robots-Tag header is the right tool, because a PDF has no head section for a meta tag. A header carrying the Bingbot training preference applies the same rule to the document that the meta tag applies to a page. And for everything at once, the site-wide robots.txt rules remain the broadest net: a crawler that cannot fetch the file cannot train on it, whatever format it is in.
The EU angle: machine-readable reservation
If your audience or your legal exposure is in Europe, there is one more framework worth knowing. The EU's copyright rules allow rights holders to reserve their works from text and data mining by machine-readable means, and the IPTC's Generative AI Opt-Out Best Practices — updated in March 2026 — recommend a layered approach: a visible rights-reservation statement on the page, the reservation in metadata, and a site-wide declaration file where appropriate. The IPTC guidance specifically discusses the noai and noimageai directives, the noarchive tag, and Google's nosnippet behaviour, and it recommends using noarchive plus noai and noimageai together for search-facing content. That recommendation predates Microsoft's March 2027 deadline for the old tags, so treat the IPTC document as the legal framing and this guide's timetable as the current technical reality. The two agree on the principle: state the reservation in more than one machine-readable place, so that no single ignored signal leaves you with nothing.
For the site-wide declaration, the emerging standard is a TDM reservation file — the machine-readable statement of which parts of a site may be used for training. It is still early, adoption is thin, and it is not a substitute for robots.txt today. But it is the direction European regulators point to, and it is worth watching alongside the shared training token. Our guide to AI crawlers and EU copyright covers the TDM opt-out in depth.
Frequently asked questions
Do noai and noimageai meta tags actually work?
Only in a limited sense. They were introduced by DeviantArt in 2022 and are recognised by some smaller platforms and dataset tools, but Google and Bing do not list them among their supported robots directives, and the major AI companies document control through robots.txt instead. Treat them as a courtesy statement, not a control, and back them with robots.txt rules and Bing's documented per-page tags.
What is the difference between noai and noimageai?
noai asks that a page's content not be used to train AI systems; noimageai asks the same specifically for images. Both are informal directives from 2022, neither is part of a web standard, and both share the same limitation: support depends on the operator choosing to honour them.
Does Google support any page-level AI opt-out?
Yes, in a narrow sense. Google documents that nosnippet prevents content from being used as a direct input for AI Overviews and AI Mode, as well as removing it from search snippets. For training, Google's documented control is the Google-Extended token in robots.txt. Google does not document support for noai or noimageai.
What are Bing's ai-training and ai-grounding meta tags?
They are per-page preferences Microsoft introduced in October 2026 to replace noarchive and nocache as AI signals. ai-training:disallow opts a page out of model training; ai-grounding:disallow opts it out of being used to ground Microsoft's AI answers. They are set with a Bingbot meta tag or an HTTP header, they override the site-wide robots.txt setting, and Bing starts reading them on 1 February 2027.
Do I still need robots.txt if I use meta tags?
Yes. Meta tags work page by page and only on HTML pages; robots.txt states the preference for the whole site, covers every file type, and is read before any page is fetched. The two layers are complementary: robots.txt for the default, meta tags and headers for the exceptions. Our llms.txt vs robots.txt guide explains how the file-based controls fit together.
When do noarchive and nocache stop working as AI controls?
For Bing, on 1 March 2027. After that date they revert to being ordinary search directives. If your site relies on them today, move to the shared training token in robots.txt and Bing's new per-page tags before the deadline, and check Bing Webmaster Tools on 15 December 2026 for the temporary grounding preference.
The honest summary: noai and noimageai are a polite note the biggest AI companies never agreed to read. The controls that count are the shared training opt-out in robots.txt, Bing's per-page ai-training and ai-grounding tags, and nosnippet where verbatim quoting is the concern. Add the old tags if you like — just don't audit your protection by them.
