How to write meta descriptions for AI citationsEntity first. Value second. 150–160 characters.
Meta descriptions still matter — AI engines lift them into source-card displays much as Google's SERP does. The right pattern is entity-first: name the subject of the page in the first phrase, then say what value the page delivers. This page covers length, format, and the specific patterns that work across all four major AI engines.
By Martin Yarnold · UpdatedThe pattern in one example
Good
"Promagen Sentinel is a B2B AI visibility audit service that measures whether AI engines (ChatGPT, Claude, Perplexity, Gemini) can find, read, and cite your site."Entity (Promagen Sentinel) first, category (B2B AI visibility audit service) second, value (measures...) third. ~155 characters.
Avoid
"Worried AI engines can't see your site? Get the audit that's helping leading B2B companies show up in ChatGPT, Claude, Perplexity, and Gemini answers — try it today!"Hook-first, entity buried, multiple CTAs in one description. Engines struggle to extract the subject.
Frequently asked questions
What length should a meta description be for AI citations?
150–160 characters works across both classic SEO (Google's SERP truncation point) and AI answer cards. Going longer is not penalised but the back half is rarely rendered; going shorter under-utilises the slot. The number is the same as SEO best practice — meta description length has not changed for AI engines because the underlying HTML element has not.
Should I lead with the entity (company, product, concept) or the value proposition?
Entity first, value second. AI engines need to disambiguate the page's primary subject quickly; an entity-first description tells the engine "this page is about X" before "X helps you do Y". The pattern "<Entity> is a/the <category> that <does Y>" works reliably. Avoid descriptions that lead with a hook or rhetorical question — they read well to humans but engines struggle to extract the entity.
Do keywords still matter in meta descriptions for AI engines?
Less than for traditional SEO. AI engines extract semantic meaning, not keyword density — stuffing the description with target keywords hurts more than it helps because it reads unnaturally and engines penalise low-quality patterns. The right approach: include the entity name once, the primary concept once, and write naturally. The keyword you would have stuffed will appear organically in good copy anyway.
Should I rewrite every existing meta description?
Prioritise the top 20% of pages by traffic and conversion. Rewriting 500 meta descriptions for a long-tail tail is rarely a good use of time. The top 20% drive most engine retrieval and most click value; rewriting those with the entity-first pattern produces measurable lift. Long-tail pages benefit more from getting metadata populated at all than from rewrite optimisation.
Do AI engines use the meta description differently than Google's SERP?
Less than you would think. AI engines tend to lift the meta description into source-card displays similarly to how Google's SERP uses it — as a concise summary of the page below the title. The exception: some engines blend the meta description with extracted page content for the answer surface, which means a meta description that lies about the page (a common SEO sin) is more visibly punished by AI engines than by Google.
Should product pages and content pages follow the same pattern?
Same skeleton, different priorities. Product pages: lead with the product name, follow with the primary use case and price/availability if stable. Content pages (guides, definitions, comparisons): lead with the concept, follow with what the page teaches. The shared rule is entity-first, intent-second. The difference is which entity (product vs concept) and which intent (purchase vs learning).