How to write meta descriptions for AI models, not just Google

For fifteen years, meta descriptions had one job: generate clicks in Google search results. You wrote 155 characters, tested them for click-through rate, and moved on. The tag was a marketing asset, not an information asset.

That changed when AI models started crawling, indexing, and citing web pages. Perplexity, ChatGPT with browsing, Google AI Overviews, and Claude all consume the full HTML of every page they process. The meta description tag is one of the first things they read, because it is the most compressed summary of what the page claims.

Before the model processes your 2,000-word article, it has already read your 300-character description. If that description is vague, the model starts with a vague understanding of your page. If it is specific, the model knows exactly what your page argues and whether it is worth citing.

The practical consequence: a meta description written for Google alone leaves value on the table for AI. A meta description written for AI alone may display poorly in SERPs. You need a structure that serves both, and the structure is not complicated once you understand what each audience extracts.

Google uses your meta description as one input when generating the snippet that appears below your title in search results. It does not use the tag as a ranking factor, but it does read it. Three things to know about how Google handles it.

Google truncates the display. The SERP snippet on desktop shows approximately 920 pixels of text, which translates to roughly 150-160 characters depending on character width. Narrow characters like “i” and “l” buy you more; wide characters like “m” and “w” cost you. On mobile, the display is narrower: roughly 120 characters.

The “155 characters” guideline is an approximation because Google uses pixel width, not character count. A description of 165 narrow characters may display fully. A description of 145 wide characters may get truncated. For practical purposes, treat 150 characters as the safe zone for full display on both desktop and mobile.

Google often rewrites your snippet. Google sometimes generates its own snippet from your page content rather than using the meta description you wrote. It does this when it believes page content better matches the specific search query.

This does not mean the meta description is wasted. AI models always use the tag as written, because they do not generate alternative snippets. The rewrite problem is Google-only. And the characters beyond 150 are where AI extraction happens.

AI models that crawl or process web pages read the full HTML source. There is no truncation. The meta description tag is consumed alongside the title tag, heading structure, body content, structured data, and OpenGraph tags. But the meta description occupies a specific role in how models process a page.

It is the first summary the model encounters. Before processing 2,000 words of body content, the model reads a 300-character description that tells it what the page claims, what evidence it offers, and what position it takes. A vague description (“Learn about cybersecurity solutions”) gives the model nothing to anchor on. A specific description (“Enterprise cybersecurity platforms are evaluated on three criteria AI models check first: detection coverage, false positive rates, and containment speed”) tells the model exactly what this page contributes to a buyer’s question.

It signals whether the page takes a position. AI models building buying recommendations need sources that commit to a claim. A meta description that hedges (“We explore the pros and cons”) signals fence-sitting. A meta description that commits (“Three criteria differentiate providers, and most vendors address only one”) signals a citable source.

The model has not yet read the body content, but the description has already influenced whether the page is worth reading in full.

It has no character limit. The HTML meta description tag accepts any length. There is no technical truncation. AI models read as much as you write.

The practical limit is usefulness: beyond 320 characters, a description starts repeating what the body content already covers. The sweet spot is 280-320 characters, enough to include the claim, the evidence type, and the specificity that distinguishes your page from every other page on the same topic.

The technique is front-loading. Write the meta description in two layers. The first layer is a complete summary that works within Google’s display window. The second layer adds the specificity that AI models need. Both layers live in the same tag, and the structure is the same every time.

Layer 1 (characters 1-150): write one to two sentences that contain the page’s main claim and target reader. No filler, no hedging, no “in this article we explore.” State the position. If the page argues that most vendors address only one of three criteria buyers care about, say that in the first sentence. The reader (and Google) should understand what this page does in under 150 characters.

Layer 2 (characters 150-320): add the differentiating specificity. Data points (“from 63 buyer scenarios”), named markets (“across the UK cybersecurity market”), content scope (“six sections covering criteria alignment, elimination triggers, and switching dynamics”), or the outcome (“three content opportunities almost no vendor has acted on”). This is the layer that makes an AI model choose your page over a competitor’s page on the same topic. Without it, your page is one of many results. With it, your page is the one with specific evidence.

Why 320, not 500? Because the description should compress, not repeat. If you need more than 320 characters, the body content is doing the work and the description is drifting into duplication. The description exists to tell both audiences what the page claims and why it is worth reading. It does not need to summarise the full argument.

Your page has three places where a description lives. Each one is read by a different audience, and each one should be written for that audience. Copying the same text across all three is a missed opportunity.

Writing three different descriptions for every page adds five minutes of work. The return is that each audience receives the description written for them instead of a compromise written for nobody. If you can only write one, write the meta description using the dual-purpose structure. It covers the widest audience.

The difference between a meta description that gets ignored and one that earns a citation is specificity. Here are two meta descriptions for the same hypothetical page, a guide to evaluating enterprise cybersecurity platforms.

The structure is the same regardless of topic. Layer 1: the claim and the position, within 150 characters. Layer 2: the evidence, scope, and differentiator, bringing the total to 280-320. Both layers are factual. Both layers reference the actual content of the page. The description is a summary, not a sales pitch.

Five patterns that reduce the chances of your page being cited by AI models. Each one is common, and each one has a specific fix.

1. 1

   Duplicating the title tag.

   If your title says “How to Evaluate Enterprise Cybersecurity Platforms” and your meta description says “A guide to evaluating enterprise cybersecurity platforms,” the description adds nothing. The model already read the title. Use the description to say what the title cannot: the specific evidence, the scope, the position. The title names the topic. The description names the contribution.

2. 2

   Writing for clicks instead of citations.

   Curiosity-gap descriptions (“You won’t believe what we found”) work for click-through rate. They fail for AI extraction because the model needs to know what you found before deciding whether to cite you. A specific description lets the model decide from the meta tag alone.

3. 3

   Stopping at 155 characters.

   The most common mistake and the easiest to fix. A 155-character description is optimised for Google display and nothing else. The additional 125-165 characters between 155 and 320 are free space that costs nothing. AI models read and use every character. Leaving that space empty is leaving the strongest AI signal on the table.

4. 4

   Using the same description for meta and OpenGraph.

   Frameworks and CMS platforms that auto-populate og:description from the meta description produce a compromise that serves neither audience well. The meta description is written for search context. The OG description is written for social context. Five minutes of separate writing produces two assets instead of one diluted one.

5. 5

   Keyword stuffing.

   Meta descriptions crammed with keywords read as spam to both Google and AI models. Google’s documentation explicitly warns against this. AI models interpret keyword-stuffed descriptions as low-quality signals, the same way a human reader would. Write a natural sentence that states your claim. If the sentence is specific, the relevant terms appear naturally.