Content Audit for AI Search: What to Merge, Redirect and Delete

A content audit used to be a ranking exercise. You pulled every URL, looked at sessions and positions, refreshed the ones slipping, and left the rest alone because pages were free to keep. That logic held while search rewarded coverage. It does not hold now. AI engines retrieve far more pages than they ever show a reader, which means being findable stopped being the win and being selected became the only thing that counts.

The number that reframes the whole exercise comes from an AirOps analysis reported by Search Engine Land in March 2026. Across 548,534 pages retrieved during 15,000 prompts, ChatGPT cited only 15% of what it pulled. The other 85% was read, weighed, and dropped before the answer was written.

So the question your archive poses is no longer “does this page rank.” It is “when an engine pulls nine of my pages on the same subject, which one wins, and what are the other eight doing besides splitting the vote.” This is the audit for that question: the criteria, the decision tree, the traffic objection, and the part nobody writes about, which is getting sign-off to remove work somebody in your building was proud of.

What is a content audit, and what changed about it in 2026?

A content audit is a full inventory of every page you have published, scored against a decision: keep, improve, merge, redirect, or remove. What changed in 2026 is the scoring criterion. The old one asked whether a page still earned traffic. The new one asks whether a page still earns a citation, or whether it is quietly competing with a stronger page of your own for the same one.

The demand for this is not niche. In Semrush’s US database as of August 2026, “content audit” runs 4,400 searches a month at a $8.97 CPC, and “what is a content audit” pulls 22,200. The narrower operational terms are much smaller: “content pruning” at 390, “content decay” at 320, “content consolidation” at 50. Read that spread honestly. Plenty of people want to be told how to look at their archive. Very few are searching for permission to cut it.

Two mechanics make cutting matter more than it used to. First, retrieval is cheap and selection is scarce, per the AirOps figures above. An engine will happily pull all nine of your overlapping posts, then cite one. Second, the pages that win selection skew heavily toward pages that already rank: 55.8% of cited pages sat in Google’s top 20, and position-1 pages were cited 3.5 times more often than pages outside the top 20. Anything that dilutes the ranking strength of your best page on a subject is now costing you twice.

That is the argument for consolidation in one line. Nine mediocre pages on one head term do not give you nine chances at a citation. They give you one chance, taken by whichever of the nine ranks least badly.

Does deleting old blog posts actually help SEO?

Not by itself, and the most useful correction on record comes from Google. In August 2023, Gizmodo reported that CNET had deleted thousands of old articles, describing the practice internally as content pruning that “sends a signal to Google that says CNET is fresh, relevant and worthy of being placed higher than our competitors.” Danny Sullivan, Google’s Public Liaison for Search, responded publicly and bluntly: “Are you deleting content from your site because you somehow believe Google doesn’t like ‘old’ content? That’s not a thing! Our guidance doesn’t encourage this.”

Sullivan went further, and the nuance in the follow-up is the part worth keeping. An outdated page “isn’t likely to rank well,” he wrote. “Removing it might mean, if you have a massive site, that we’re better able to crawl other content on the site. But it doesn’t mean we go, ‘Oh, now the whole site is so much better’ because of what happens with an individual page.”

Hold both halves of that. Deletion is not a ranking lever, but it can be a crawl and clarity lever on a large site, and it is often the honest end state for a page nobody should read. SEO practitioners quoted in the same reporting agreed with the caveat rather than the ban: Lily Ray of Amsive noted that the claim deletion never helps “isn’t always true,” and Chris Rodgers of CSP framed pruning as an advanced practice where updating or redirecting comes first and deleting without a redirect is the last resort.

So the correct framing for your audit is not “delete old posts.” It is “stop running eleven pages where one belongs, and stop indexing pages you would be embarrassed to have quoted.” Age is not the variable. Redundancy and quality are.

How do you decide what to merge, redirect, or delete?

Run every URL through five questions in order, and stop at the first one that gives you an answer. The order matters, because the expensive mistakes come from jumping straight to the traffic column.

1. Does anything link to it? Referring domains outrank everything else in this decision. A page with real external links is never a delete. It is a merge target, a redirect, or a rewrite. Losing earned links to save a URL is the one genuinely irreversible mistake in this process.
2. Does another page of ours answer the same question? If yes, you have found a cannibalization pair, and the choice is between them, not about them. Keep the stronger page, fold the unique substance of the weaker one into it, redirect the weaker URL to the stronger.
3. Does it bring people who do something? Separate qualified traffic from a number. A post with 2,000 monthly sessions and zero downstream action is worth less than a post with 60 sessions that produces two conversations a quarter. Sessions alone have been lying to content teams for a decade.
4. Is it still true? A factually stale page is worse than an absent one now, because engines will quote it. Wrong dates, retired products, superseded pricing and dead screenshots are all citation risks, not just reader problems.
5. Would we be comfortable seeing it quoted? This is the AI-era version of the quality question, and it is the one that finally lets people say no out loud. If you would not want a passage from the page read back to a prospect as your position, either fix it or remove it.

The actions those questions produce map cleanly. Keep and leave alone is the default and should cover most of the archive. Improve applies to pages that rank, convert, or define your position and have simply drifted, which is the refresh workflow we covered separately in why AI engines cite newer pages. Merge applies to overlapping pages on one head term. Redirect applies to any merged or removed URL with links, traffic, or a sensible destination. Delete without a redirect applies only to pages with no links, no traffic, no sensible parent, and no reason to exist, and that list is usually short.

One practical note on merging. Do not paste four posts together. Take the genuinely distinct sections from each, write one page that answers the head question directly at the top, and throw away the transitional filler that made each original post feel like an article. The merged page should be shorter than the sum of its parts and denser per paragraph.

How do you audit content for answer engine readiness?

Add one column to your existing audit: for each surviving page, name the specific question it is the best answer to, and check whether that answer is stated plainly near the top. Pages that cannot pass that test are not audit failures, they are rewrite candidates, and this is where the AI-era audit diverges from the classic one.

The AirOps data gives the column some teeth, because citation rates vary sharply by what kind of question is being asked. Product-discovery prompts produced citations 18.3% of the time and how-to prompts 16.9%, while comparison prompts came in at 13.1% and validation prompts, where someone is checking a claim they already half-believe, at just 11.3%. If your archive is mostly validation-flavored thought leadership, it is competing in the hardest citation category on the board.

Two more findings from the same dataset should change what you keep. Fan-out is now a second citation surface: 89.6% of prompts triggered two or more follow-up searches, those 15,000 prompts expanded into 43,233 queries, and 32.9% of cited pages appeared only in fan-out results rather than against the original prompt. Then the detail that ought to stop a few deletions cold: 95% of those fan-out queries had zero traditional search volume.

Which means a page can be worth keeping despite a flat traffic line, because the sub-question it answers well never appears in a keyword tool at all. Before you cut a low-traffic page, check whether it is the only place you answer a specific, narrow, obviously real question. That is a fan-out asset, and your analytics will never tell you so. Finding those holes deliberately is the other half of this work, which we walk through in content gap analysis across the user journey.

Will traffic drop after you prune, and how do you defend it?

Expect a dip in raw page count and often in raw sessions, and be clear before you start about which sessions you are willing to lose. This is the objection that kills most audits, and it is usually argued with the wrong number.

Start by separating the two things a removed page was doing. If it brought people who read one page and left, you are losing a metric. If it brought people who converted, subscribed, or turned up in a sales conversation, you are losing revenue, and that page should not have been on the list. Agree the distinction in writing with whoever owns the traffic target, before the first redirect ships. The version of this conversation that goes badly is the one held six weeks after the fact.

Two mechanical safeguards keep the dip small. Redirect rather than remove wherever a sensible destination exists, so link equity and any residual demand land somewhere. And stage the work in batches by topic cluster, because cutting 400 URLs in one release leaves you no way to read the result. Set expectations on timing while you are at it: recrawling and re-evaluation are not instant, and a fair internal promise is that nobody draws conclusions for a full quarter.

Then there is the risk nobody budgets for, which is volatility you do not control. Semrush tracked more than 100 million AI citations across 230,000 prompts over 13 weeks in 2025 and watched ChatGPT’s citation rate for Reddit fall from close to 60% of responses in early August to roughly 10% by mid-September, with Wikipedia sliding from about 55% to under 20% in the same window. Neither site changed. The engine did. Keep that in view when you attribute a post-audit movement to your own work, and keep your baseline measurements honest enough to tell the two apart.

How often should you run a content audit, and how long does it take?

Run a full audit once a year and a light one quarterly, and budget the full pass in weeks rather than days. Both of those questions show real search demand in Semrush’s US data, at 20 searches a month each for “how often should you do a content audit” and “how long does a content audit take,” which tells you people are trying to scope this before they commit to it.

A workable cadence for a mid-market site with a few hundred posts:

1. Annual full pass. Every URL through the five questions. For 300 to 500 posts, expect two to four weeks of focused work including the rewrites, not counting the approval cycle.
2. Quarterly light pass. Only pages that moved: new cannibalization pairs, factual staleness, anything that lost or gained sharply. An afternoon, not a project.
3. Publish-time discipline. Before any new post goes live, check whether it should be a section of an existing page instead. This is the single highest-return habit here, because it stops the archive from regenerating the problem you just spent a month fixing.

Two edge cases deserve honesty. A site under a few dozen pages does not have a pruning problem and should spend the same hours improving what it has. And a site whose archive is largely news or dated announcements is not a consolidation candidate at all, because those pages are not competing for one evergreen answer. They are a record, and Sullivan’s point about old content applies to them directly.

The organizational part is the reason audits stall rather than fail. Somebody wrote those posts. Name the criteria before you name any URLs, get the criteria agreed, then run pages through them in public. When the standard is visible and the decision is mechanical, the conversation stops being about anyone’s work and starts being about the rule. That is usually enough.

This is content marketing maintenance work, and it is the leg most teams skip. Finding gaps is exciting, refreshing is easy to justify, and deciding what to remove is the one that requires a standard and a spine. It is also the one that decides whether your best page gets to win.