The Content Audit That Actually Matters for AI Search: Why Merging Thin Pages Beats Adding More

The Diluted Authority Problem

Run enough live citation experiments across ChatGPT, and a pattern surfaces that most SEO teams never account for: AI engines don’t reward the site with the most pages on a topic. They reward the page with the clearest, most concentrated signal of authority.

We saw this directly while tearing down how ChatGPT handles DTC eyewear brand comparisons. When we tested “best DTC eyewear brands like Warby Parker,” the model gave one brand — Ace & Tate — a full deep-dive treatment: three images, a narrative writeup, an “industry rankings” citation. Ace & Tate doesn’t even have an llms.txt file. Warby Parker, with comparable off-site volume and its own llms.txt, got folded into a comparison table instead.

The difference wasn’t technical infrastructure. It was that Ace & Tate’s off-site mentions were dense, fresh, and unambiguously positive — a single, concentrated consensus. Warby Parker’s presence, while larger, was more fragmented and included skeptical, comparative framing.

That’s the pattern worth sitting with: AI engines don’t average your content. They pick a winner. And the winner is usually whichever page (or whichever brand’s concentrated footprint) reads as the single clearest answer — not whichever entity has the most total content about the topic.

This has a direct, practical consequence for how you audit your own site. If you have three mediocre pages competing to be the “best answer” on a topic, you’re not tripling your chances of getting cited. You’re splitting the one signal that would have made a single page citable.

Why Traditional Audits Miss This

The standard content audit checklist is built for SERP logic:

• Is this page cannibalizing another page’s ranking?
• Is traffic declining?
• Does it still have backlink equity worth preserving?

These questions still matter. But they were designed for an environment where ten blue links compete for position one through ten. AI answer engines compress that down to often a single cited source, or a short table with one page singled out for elaboration.

That changes the cost of fragmentation. In classic SEO, three separate pages targeting adjacent long-tail terms might all rank quietly on page two and collectively bring in decent traffic. In AI search, three separate pages splitting the same core answer usually means none of them reads as authoritative enough to be the one the model expands on — the model has no way to tell you’re actually the deepest resource here, because that depth is scattered.

The fix isn’t more content. It’s consolidation with intent.

The Five-Category Framework, Rebuilt for AI Visibility

The classic Keep / Improve / Merge / Redirect / Remove structure still holds up — but each category now has a second job: managing your site’s citation signal, not just its ranking signal.

Keep

The page already reads as a clear, singular answer to its target query. Structure and depth are right. Action here is light-touch: refresh statistics, update dates, confirm the page still reflects current reality. Don’t restructure something that’s already unambiguous — restructuring for its own sake is how you accidentally re-fragment a signal that was working.

Improve

This is where most money pages land. The page is targeting the right query and sitting on the right URL, but the content is thin, outdated, missing schema, or short on the kind of specific, verifiable detail (exact pricing, named policies, sourced claims) that both users and AI crawlers treat as trust signals. Keep and Improve aren’t competing categories — they’re sequential. You keep the URL and its accumulated authority, then improve what’s actually on it.

The AI-citation angle here: our llms.txt findings showed that having the file isn’t what determines citation quality — coverage is. Glossier’s llms.txt explicitly covered returns, refunds, exchanges, and international policy in detail, and got a single clean citation straight to its own domain. Warby Parker had an llms.txt file too, but it didn’t cover returns — so the model patched the gap with third-party aggregators instead. “Improve” means asking not just “is this page good” but “does this page actually answer the specific sub-questions someone — or some model — would ask next.”

Merge

This is the category that does the most work for AI visibility, and the one teams most often skip or do incompletely.

If two or more pages are competing for the same intent — near-identical H1s, overlapping title tags, answering the same underlying question from slightly different angles — don’t leave them to compete. Identify the strongest candidate (most backlinks, most historical traffic, best current ranking), fold every unique, valuable piece of the weaker pages into it, and retire the rest.

The critical step people miss: merge without a 301 redirect just leaves orphaned indexed pages competing with your consolidated one. You haven’t actually concentrated the signal — you’ve just added a fourth page to the pile. A merge isn’t complete until the losing URLs point to the winner.

Redirect

Some pages carry real equity — backlinks, occasional traffic, brand history — but no longer deserve to exist as a standalone destination: a discontinued service, a renamed product line, an old campaign page. These get a 301 to the most relevant live page. You’re not removing value, you’re relocating it.

Remove (410)

Reserve this for pages with genuinely zero remaining value: no backlinks, no traffic, no equity to pass forward, content that’s fully obsolete (expired events, old job listings, auto-generated duplicate pages). A 410 tells crawlers — including AI crawlers like GPTBot and ClaudeBot — “this is gone on purpose, stop checking.” That’s a faster, cleaner signal than a 404, which reads as ambiguous or temporary.

The Redirect vs. 410 Mistake Worth Avoiding

Before sending anything to 410, check it in GSC or your crawler of choice for backlinks and residual traffic. If a page has either, it’s a 301 candidate, not a 410 candidate — 410-ing a page with real equity just throws that equity away instead of redirecting it somewhere useful. This is a five-minute check that prevents an easily avoidable, permanent mistake.

A Worked Example: Technical SEO Content Clusters

Imagine a site running separate posts on “robots.txt directives for AI crawlers,” “setting up llms.txt,” and “AI crawler access configuration.” On paper, these look like three distinct topics. In practice, they’re often answering the same underlying question from three angles: how do I make my site legible to AI crawlers?

Rather than letting all three compete quietly, the audit move is:

1. Identify which one has the strongest current signal (traffic, backlinks, rank).
2. Fold the genuinely unique value from the other two into it — the specific llms.txt coverage nuances, the specific crawler directive syntax, whatever isn’t redundant.
3. 301 the other two into the winner.
4. Result: one deep, well-linked, comprehensive resource — the kind of single page that reads as the obvious answer rather than three pages that each read as partial.

That’s the difference between adding content and concentrating authority.

Audits Are an AEO Practice, Not Just Housekeeping

Content audits get treated as cleanup — something you do once a year to tidy the sitemap. That undersells what they actually do. Every merge decision, every redirect, every 410 is a decision about how concentrated your site’s authority signal is on a given topic. Do it well, and you’re not just cleaning house — you’re engineering the conditions under which an AI engine can look at your site and find one clear, citable answer instead of three ambiguous ones.

This sits alongside the other pieces in the same thread: what determines whether AI engines trust a page enough to cite it. The Reddit Paradox looked at community consensus as a signal. The llms.txt piece looked at crawlability and coverage. This one is about consolidation — making sure the signal you’ve built isn’t scattered across pages that are quietly competing with each other instead of standing together.