AI prompts for SEO work

Where AI genuinely helps in an SEO workflow, and where it does not

SEO work splits into research, drafting and verification, and AI's usefulness varies sharply across the three. Drafting is where it earns its keep: turning a cluster of related keywords into a content brief, writing title tag variants, or restructuring a page's headings against a target structure are all tasks a well-prompted model does quickly and well.

Research is where the limits show. A model has no live connection to Google's index, no current ranking data, and no real search-volume numbers unless a tool feeds them in. Asked "what is the search volume for electric van leasing UK," a model without live data access will produce a plausible number with total confidence, and that number is invention. Google's own Search Central documentation is explicit that AI use itself does not violate spam policies — the policy targets mass-produced unhelpful content, not the production method — which means the responsibility for accuracy sits entirely with the human running the workflow, tool or no tool.

Verification is the third stage and the one most often skipped. A brief generated from real ranking pages still needs a person to check that the pages cited are actually ranking today, because rankings move and a stale citation ages badly inside a live prompt library.

Keyword clustering: group by intent, not by shared words

The classic mistake in AI-assisted clustering is asking the model to group keywords that "look similar," which produces clusters based on shared vocabulary rather than shared intent. "Van lease calculator," "van leasing cost," and "how much does it cost to lease a van" share almost no words in common with "lease vs buy van" yet target adjacent stages of the same buyer journey — a model told to cluster by intent catches this; one told to cluster by similarity misses it.

The instruction to flag ambiguity matters because forcing every keyword into a clean bucket produces false confidence. A keyword like "van leasing" alone, with no modifier, genuinely sits between informational and commercial intent, and a good cluster output says so rather than picking one arbitrarily.

Content briefs: feed it the real SERP, not a description of the topic

A content brief written from a topic description alone produces generic structure — introduction, three body sections, conclusion — because that is the statistical shape of articles on most subjects. A content brief written from the actual top-ranking pages' headings produces something closer to competitive intelligence.

This prompt cannot run without the actual headings pasted in, which is the point: it converts a request for creative writing into a request for pattern analysis on real data, and pattern analysis is a task language models handle well.

Title tags: prompt for pixel width, not character count

Google truncates title tags based on approximate pixel width, not a fixed character count, because wide characters like "W" and "M" consume more horizontal space than narrow ones like "i" and "l." A title using mostly wide characters can truncate well under 60 characters, while one using mostly narrow characters can run past 60 and still display fully. The commonly cited working target is around 580 pixels, which maps loosely to 50-60 characters depending on the specific letters used.

After generating variants, paste each into a SERP snippet preview tool to confirm it does not truncate, since no prompt-based pixel estimate substitutes for an actual rendering check.

Meta descriptions that earn the click without inventing a promise

A meta description does not affect ranking directly, but it affects click-through rate, which is an observable, measurable signal. The failure mode in AI-generated descriptions is overclaiming — promising "the ultimate guide" or "everything you need to know" when the page delivers something narrower, which increases bounce rate on the very traffic the description attracted.

Requiring the model to work from a list of what the page actually covers, rather than the topic in the abstract, is what prevents the "ultimate guide" problem — the model cannot promise comprehensiveness it was never told exists.

Internal linking: ask for anchor text variation, not just placement

A common AI-generated internal linking failure is anchor text repetition — using the identical phrase "click here to learn more" or the identical exact-match keyword across dozens of links. Google's guidance on link best practices recommends descriptive, varied anchor text over generic or repeated phrases, both for user experience and because heavily repeated exact-match anchors can look manipulative.

Restricting the model to a supplied list of real pages prevents it from suggesting links to URLs that do not exist — a common failure when the model is simply asked to "add internal links" without a source list.

Schema markup: generate the JSON-LD, then validate it

Structured data generation is one of the more reliable AI-assisted SEO tasks because JSON-LD has a strict, well-documented syntax the model has seen extensively in training, and the output is either valid or it is not — there is no ambiguity to hide behind.

Always validate the output afterward with Google's Rich Results Test or the Schema.org validator rather than trusting that generated JSON-LD is correct by inspection — a single missing comma or mismatched bracket will fail silently in a manual read and fail loudly in the validator.

Technical SEO audits: AI reads the crawl data, it does not replace crawling it

A model cannot crawl your site. It can, however, read the output of a crawl — a Screaming Frog export, a Search Console coverage report, a log file sample — and summarise patterns a person would otherwise scan manually for an hour.

This use case is genuinely strong because it plays to a real strength — finding patterns across a large, structured dataset — without asking the model to know anything it was not given.

Writing for AI answer engines is the same discipline as writing for Google

2026 passage-retrieval research on how AI answer engines source their responses found that cited passages typically run 74 to 148 words, and that dense, fact-heavy paragraphs — those naming 3 or more entities or figures per 100 words — get pulled into answers substantially more often than vaguer prose, by a factor of roughly 2.4. The Princeton GEO study separately found statistics moved AI visibility by up to 41%, the single strongest lever measured, while keyword stuffing produced no benefit.

This is not a separate skill from traditional SEO writing — it is the same discipline of specific, sourced, well-structured prose, applied with slightly more attention to whether any single paragraph could stand alone if an answer engine extracted it.

FAQ sections: only write questions people actually ask

An FAQ section built from invented questions rarely earns FAQPage rich results engagement, because the questions do not match real search behaviour. The stronger input is a list of questions people actually typed, pulled from Google's "People also ask" boxes, from Search Console's query report, or from a tool like AlsoAsked.

Restricting the model to real, sourced questions is what makes the resulting FAQPage schema genuinely useful rather than a set of plausible-sounding questions nobody searches for.

E-E-A-T signals: AI can draft them, it cannot manufacture credibility

Google's Search Quality Rater Guidelines describe experience, expertise, authoritativeness and trust as factors human raters weigh when assessing content quality, particularly for topics that could affect a reader's health, finances, or safety — the category Google labels Your Money or Your Life. A model can help surface where a page is missing these signals; it cannot substitute a real author's credentials for a missing byline, and it should never be asked to invent one.

The prompt is deliberately constrained to flagging gaps rather than filling them, because E-E-A-T signals that are fabricated are not merely unhelpful — they are the exact failure the guidelines exist to catch, and a page caught doing this loses more trust than one that simply lacks a byline.

A complete worked example: brief to published draft

Combining several of the prompts above into one workflow, for a page targeting "van leasing calculator UK":

1. Cluster: Paste a 40-keyword export including volume data. Ask for intent clusters. Result: a commercial cluster of 12 keywords centred on cost comparison.
2. Brief: Paste the H2 headings from the top 5 ranking pages. Ask for required topics plus one gap. Result: all 5 competitors cover monthly cost; none show a lease-versus-buy break-even calculation.
3. Draft: Using the brief, write the article with the six-part prompt structure — task, reader, source material (your actual lease terms), format, constraints, acceptance criteria.
4. Title and meta: Generate 5 title variants at 50-55 characters and 3 meta description variants naming what the page actually covers.
5. Schema: Generate FAQPage JSON-LD from 6 real questions pulled from Search Console, then validate in Google's Rich Results Test.

Each step supplies real data the previous step produced or the business already had — at no point does the model invent search volume, ranking data, or a competitor's content. That discipline is what separates AI-assisted SEO from AI-generated guessing.

What to verify after any AI-assisted SEO prompt

Four checks apply regardless of which prompt above produced the draft. First, confirm every statistic traces to a real, dated source — an AI-generated SEO article citing an invented statistic is the fastest way to publish a page that damages credibility rather than building it. Second, check that any competitor or ranking-page claim reflects the current SERP, since a model working from headings pasted three weeks ago may describe a page that has since changed.

Third, validate any generated schema markup in an actual validator, not by reading it. Fourth, run the finished piece past Google's own guidance on helpful content — the question is not whether AI was used, but whether the page would satisfy a person who searched for it, which is the standard Google's Search Central documentation states it evaluates against regardless of production method.