AI Search Content Audit: Keep, Merge, Update, Remove

Most AI-search content plans open with "what should we publish?" On an established site, the sharper question is "which existing URL should own this answer?"

Publish another variation and you split internal links, confuse the preferred canonical, repeat claims without adding evidence and make measurement noisier. Delete weak pages indiscriminately and the danger is equal: a modest URL may serve a distinct audience, earn links, support a buyer journey or provide the exact passage an answer engine needs.

An AI search content audit should therefore make four evidence-led decisions: keep, merge, update or remove. This guide supplies the scoring model, evidence pack and implementation controls.

Core principle: audit by user job and evidence contribution, not by keyword alone. Two pages can use different terms but answer the same question; two pages can share a keyword while serving genuinely different decisions.

TL;DR: An AI search content audit scores every existing URL by user job and evidence contribution, then makes one of four evidence-led decisions: keep, merge, update or remove.

Key Takeaways

  • Audit by user job and evidence contribution, not by keyword: two pages can use different terms but answer the same question, and two can share a keyword yet serve genuinely different decisions.
  • Every URL earns one of four decisions: keep (distinct job still satisfied), merge (same job, one stronger destination), update (valid job, weak execution) or remove (no recoverable purpose).
  • Score contribution on 100 points across seven dimensions, but never automate the final action from the total: a page scoring 25 can hold the cluster's only primary interview; one scoring 80 can duplicate a stronger canonical.
  • Google's guidance says SEO fundamentals still apply, no special AI schema or format is required, and warns against many pages for query variations; citation metrics from Bing's AI Performance reporting are observational, not a ranking formula.
  • Work in reversible batches on one cluster, capture a baseline, ship one controlled release, then review at 30, 60 and 90 days: these are review points, not guaranteed stabilisation periods.
Content cards being sorted during an AI search content audit

What has actually changed, and what has not?

Established SEO fundamentals still apply. Google's current official guidance for generative Search recommends unique, non-commodity content, says no special AI schema or content format is required, and warns against creating many pages for query variations. Google also documents canonicalisation as selecting a representative URL from duplicate or similar pages.

These points do not prove that merging pages will produce AI citations. They do establish a defensible operating direction: create clear ownership, consolidate duplication and invest in distinctive evidence rather than manufacturing surface variation.

Bing's AI Performance reporting adds another useful evidence class, citations and cited URLs, but citation metrics are observational. They do not reveal a universal ranking formula or prove downstream revenue. SAGEO therefore measures search, answer and generative visibility separately from commercial outcomes. If that separation is new to you, see how SAGEO works.

Why should you build clusters from pages, not keywords?

Because a user job, not a keyword, is what a page must own. Export indexable URLs with title, canonical, status code, word count, last substantive update, organic queries, clicks, links, conversions and any available AI-citation observations. Then group pages around a user job:

  • learn a concept;
  • compare approaches;
  • diagnose a problem;
  • implement a task;
  • evaluate a supplier; or
  • buy a service.

A cluster called "AI visibility" is too broad. Better clusters include "how to measure AI citations", "how to select an AI search agency" and "why AI referral traffic disagrees with citation growth". These can link to one another without competing for the same answer.

Use embeddings or language models to suggest similarity, but require human review. Semantic proximity is a lead, not a deletion decision.

What goes in the evidence pack for each URL?

Twelve fields, one row per URL. Missing data is not zero; mark it unknown, so a page with no configured conversion event is not scored as commercially worthless.

  1. intended audience;
  2. user job and funnel stage;
  3. primary answer in one sentence;
  4. unique first-hand evidence;
  5. sources and update date;
  6. organic queries and trend;
  7. backlinks and referring domains;
  8. conversions or assisted outcomes;
  9. internal-link role;
  10. canonical/indexation state;
  11. observed AI citations/mentions by stable prompt set;
  12. nearest overlapping URL.

How do you score a page's real contribution?

Score contribution, not vanity, on a 100-point decision scale across seven dimensions:

DimensionWeightFull marks mean
Distinct user job20serves a clearly different decision
Original evidence20first-hand data, expert method or examples
Search demand15sustained relevant impressions/clicks
Authority15useful links, references or internal role
Commercial value15qualified outcomes or clear assisted role
Citation utility10observed citations or highly extractable evidence
Freshness/accuracy5current claims and maintained sources
Overlapping document streams consolidating into one authoritative web page

Do not automate the final action from the total. A page scoring 25 can still contain the only primary interview in the cluster; a page scoring 80 can still duplicate a stronger canonical. The dimensions reveal why a decision is being considered, not the decision itself.

When should you keep a URL?

Keep a URL when it has a distinct job and still satisfies it. It need not rank first or earn AI citations today.

Good keep signals:

  • different audience, geography or decision stage;
  • unique first-hand evidence;
  • strong links or qualified outcomes;
  • a necessary reference or product-support role;
  • clear separation from neighbouring pages.

A keep decision may still require title, internal-link or schema hygiene. "Keep" means preserve its independent purpose, not freeze it forever.

When should you merge two pages?

Merge when two or more URLs substantially answer the same user job and one stronger destination can preserve the useful evidence.

Choose the target using more than traffic:

  1. best match to durable intent;
  2. strongest relevant links and internal prominence;
  3. clearest URL and canonical state;
  4. highest-quality original evidence;
  5. best commercial continuity; and
  6. least implementation risk.

Then build a content transfer map. For each source URL, identify claims, examples, links, media and questions worth preserving. Rewrite the target as one coherent page; do not paste articles end to end. Redirect retired source URLs directly to the target, update internal links and sitemaps, and retain a change log.

Google states that redirects and canonical annotations are signals for canonicalisation, with redirects and rel=canonical described as strong signals. Do not chain redirects or leave internal links pointing at retired URLs.

When is update the right call?

Update when the user job remains valid but the execution is weak, stale or unprovable. Typical candidates have useful demand or authority but need:

  • obsolete claims replaced with current primary sources;
  • a direct answer before background;
  • first-hand evidence or named expert judgement;
  • better limitations and counterexamples;
  • clearer definitions, tables or decision steps;
  • refreshed screenshots and product facts; or
  • a meaningful bridge to the next buyer action.
Analyst reviewing clusters and traffic patterns for overlapping content

An update date should mean substantive review, not a timestamp change. Record what changed. For high-risk topics, establish an owner and review interval.

When should you remove a page entirely?

Remove when a page has no recoverable independent purpose, evidence, demand, authority or journey role, and merging would add clutter rather than value. Examples include expired announcements with no historical need, doorway pages, empty tags, thin machine-generated variations and factual pages that cannot be responsibly maintained.

Before removal, check:

  • backlinks and referral traffic;
  • conversions and assisted paths;
  • internal links;
  • legal or support obligations;
  • exact external citations; and
  • whether a relevant replacement exists.

Use a direct redirect only when the replacement genuinely satisfies the old intent. Otherwise, an appropriate 404 or 410 may be more honest. Do not redirect every removed URL to the homepage.

What is the fifth, internal-only label?

Investigate. Publicly, the framework has four actions; operationally, use investigate when evidence conflicts. For example:

  • no organic traffic, but several qualified assisted conversions;
  • strong links, but outdated and risky claims;
  • observed AI citations, but the cited passage is inaccurate;
  • similar intent, but different regulated audiences.

Give investigation an owner and deadline. It must not become a permanent parking lot.

How do you test whether two pages truly overlap?

Run a five-question interview on each pair of suspected duplicates:

  1. Would the same person search both questions in the same session?
  2. Is the recommended decision materially different?
  3. Does each page contain evidence the other cannot absorb?
  4. Would keeping both make internal linking more or less obvious?
  5. If an assistant cited one, would the other add a non-repetitive answer?

If the answers are same, no, no, less obvious and no, merging is usually the stronger hypothesis. It remains a hypothesis until monitored after implementation.

Carefully pruned plant representing evidence-led removal of weak pages

How should you implement changes safely?

Work in reversible batches; do not prune a large site in one release. Start with one coherent cluster. Capture a baseline for at least:

  • indexation and selected canonicals;
  • cluster clicks/impressions and query coverage;
  • target URL rankings;
  • crawl errors and redirect behaviour;
  • conversions and assisted outcomes;
  • stable-prompt AI citations/mentions; and
  • referral traffic from AI systems where identifiable.

Implement redirects, canonicals, internal links, sitemap changes and content in one controlled release. Validate server responses and rendered canonical tags. Then annotate analytics and search tools.

Monitor cluster-level outcomes as well as URL-level outcomes. A successful merge can reduce total indexed URLs while increasing qualified coverage. Conversely, a target gaining traffic does not compensate for lost conversions elsewhere unless the cluster result improves.

What does a 30/60/90-day acceptance plan look like?

Three review points, not guaranteed stabilisation periods:

At 30 days: verify crawling, redirects, canonical selection, indexation and broken internal links. Avoid declaring commercial success from early volatility.

At 60 days: compare query coverage, clicks and the target's passage/answer performance. Review stable prompt panels across relevant engines, without claiming causality.

At 90 days: assess qualified outcomes, assisted journeys and whether the new page ownership is operationally clearer. Decide to keep, refine or roll back specific changes.

Search and AI systems change continuously; these windows are review points, not guaranteed stabilisation periods.

What should you never do?

  • Do not delete every zero-click page.
  • Do not merge different countries or regulated audiences merely because titles resemble each other.
  • Do not use rel=canonical as a substitute for actually resolving substantially overlapping editorial pages.
  • Do not preserve weak pages because an AI tool assigned a high "citation score".
  • Do not publish a new variation before inspecting the existing cluster.
  • Do not promise citation or ranking gains from consolidation.

What should a buyer expect the audit to deliver?

More than a spreadsheet coloured red and green. Require:

Magnifying glass identifying one primary page among overlapping web content
  1. cluster map and ownership logic;
  2. page-level evidence pack;
  3. keep/merge/update/remove recommendation with rationale;
  4. transfer maps for every merge;
  5. redirect, canonical and internal-link specification;
  6. risk register and rollback plan;
  7. baseline dashboard; and
  8. 30/60/90-day acceptance criteria.

That is an implementation brief, not a generic content audit.

Your next step

If your site has years of overlapping SEO articles, do not begin an AI-search programme by adding another hundred. Audit one commercially important cluster and establish clear answer ownership first.

Request a SAGEO audit to map search, answer and generative visibility together. If you are comparing suppliers, use the AI search agency RFP scorecard to require evidence and acceptance criteria, and pair the audit with the KPI and dashboard framework to measure it.

Sources