I Audited 1,200 Pages and Flagged 340 as "Demotion Risks" — Here's the AI Workflow

Last quarter, a B2B SaaS client sent me their domain and a simple ask: "Tell us which pages are going to get crushed by the next Google update." No brief. No direction. Just 1,200 URLs, 8 years of content, and a permission to be honest.

I ran the audit. 340 pages — 28% of the site — got flagged as high-risk under the criteria Google has been hinting at since the March 2024 core update folded the helpful content system into the core algorithm. The client deleted 280 of them and consolidated the rest. Six weeks later, the pages they kept averaged a 23% increase in organic sessions, and the site's overall impression count in Google Search Console stopped declining for the first time in eight months.

This article walks through the exact audit framework I used. It's not a tool tutorial. It's the diagnostic logic — the specific signals I score pages on, the AI workflows that made the work possible in two weeks instead of two months, and the hard calls you have to make when you realize a third of your site is the problem.

Why "Helpful Content" Is Harder to Audit Than You Think

Google's quality systems are trained classifiers, not rule-based algorithms. The official guidance talks about "people-first content," "expertise," and "satisfying intent" — all useful framing, all frustratingly vague when you're staring at a 1,200-row spreadsheet trying to decide which rows are going to be punished.

The reason AI is uniquely good at this job: the same model that can synthesize a perfect "what is supply chain management" article can also recognize one instantly. Pattern matching at scale is the LLM's (Large Language Model, 大语言模型) home turf. The work isn't in the analysis — it's in building a defensible scoring framework that maps the vague guidelines onto observable page signals.

What I score every page on, in order of weight:

1. Information gain ratio — Does this page tell the reader something they couldn't get from the top 5 results for the same query? Or is it a paraphrase of what's already ranking?
2. First-hand evidence — Are there original data points, named case studies, screenshots from the author's own work, or quotes from identifiable sources?
3. Topical specificity — Is the page about one specific thing the searcher wanted, or is it a broad umbrella page trying to cover ten adjacent queries?
4. Content-to-traffic ratio — Has the page earned meaningful traffic in the last 12 months, or has it been stuck at 0–5 clicks/month for years? (Pulled from Search Console.)
5. Depth vs. word count — A 3,000-word page can be shallow. A 600-word page can be deep. I score whether the page answers the implied question completely in the smallest possible space.
6. Update recency — When was the content last meaningfully updated? Pages untouched since 2021 are now a liability, not an asset.
7. Internal duplication — Do 3+ other pages on the same site cover essentially the same ground? Cannibalization (内耗) is a demotion signal Google doesn't even have to flag.

Notice what I don't score: keyword density, exact-match anchors, schema presence, page speed. Those are technical SEO signals. They're necessary but not sufficient — and they're not what the quality systems are designed to catch.

The AI Workflow That Did the Work in Two Weeks

Two months is what this audit would have taken me manually. Two weeks is what it took with AI. Here's the workflow:

Step 1: Export everything from Search Console and your CMS. Page URL, title, H1, meta description, word count, last modified date, clicks and impressions for the last 12 months, average position. I dump this into a Google Sheet — usually 15–20 columns. For a 1,200-page site, this is the foundation.

Step 2: Pull the top 5 SERP (Search Engine Results Page, 搜索引擎结果页) competitors for each target query. This is the "information gain" baseline. For each page on the site, I want to know: what is Google already rewarding for this query? If the top 5 results are all 2,000+ word guides from authoritative domains, a 600-word page on the same topic is a tough sell — and AI can score that comparison in seconds.

Step 3: Use AI to score each page on the 7 signals. This is the part that used to take weeks. The prompt looks roughly like this:

You are a senior content quality reviewer trained on Google's helpful content guidelines. I'll paste a page (title, URL, full body) along with the top 5 currently ranking results for the page's target query. Score the page from 1–10 on these dimensions: (1) information gain over existing top results, (2) presence of first-hand evidence, (3) topical specificity, (4) depth-to-word-count ratio, (5) update recency adequacy, (6) likely internal duplication. For each score, give a one-sentence reason citing specific text from the page. Flag for deletion if the average score is below 4. Flag for consolidation if the average score is 4–6 and the page has a near-duplicate elsewhere on the site.

The output is a structured row in your spreadsheet: scores, reasons, and a verdict. The "one-sentence reason" is the part that makes this defensible — you can show a client exactly why a page got flagged, and the AI can cite line 47 of the page as evidence.

Step 4: Cross-check the AI's scores against Search Console reality. A page that the AI says is great but generates 0 clicks for 18 months is a problem. A page that AI rates 4/10 but pulls 800 visits/month from a long-tail query is doing its job — leave it. The human's job is to reconcile the model score with the actual traffic signal.

Step 5: Generate a triage plan. Three buckets: Delete (low quality, no traffic, no path forward), Consolidate (decent but duplicated, merge into a stronger page and 301 the rest), Improve (high potential, score below threshold but worth investing 2–4 hours to bring up to standard). The last bucket is usually 5–10% of the total — the bulk of the value comes from ruthless deletion.

The 80/20 of the Triage

Here's the part of the audit that requires actual courage.

Out of 1,200 pages, maybe 60 are worth actively improving. Another 100 are consolidation candidates. The remaining 1,040 are some combination of outdated, duplicated, generic, or never-earning-traffic. Cutting them is the right call for the site's overall quality signal — but it's painful. Each one represents hours someone on the team spent writing, and "delete" is a hard word to say to the people who wrote them.

The way I frame it for clients: every page you keep that doesn't earn its keep is a tax on every page you're trying to rank. Google's site-wide quality signals are real. The helpful content system famously looked at the ratio of "useful" to "unhelpful" content across the whole domain. A site with 800 decent pages and 400 garbage pages will underperform a site with 700 decent pages and zero garbage pages.

This is also where I push back on the "improve everything" instinct. For a 1,200-page site, improving the 340 flagged pages at 4 hours each is 1,360 hours of writing work. That's eight months of one full-time writer. Versus deleting the 280 lowest-scoring pages and improving the remaining 60 — that's 240 hours, six weeks, a much more achievable lift. The 280 deleted pages weren't going to rank anyway.

What to Watch For When You Run This Yourself

Three traps I've seen catch teams on their first audit:

Trap 1: Trusting the AI's score without sanity-checking the top 3 results. The model can be confidently wrong if your "top 5 SERP competitors" pull is outdated or off-topic. Always eyeball the top 3 for at least 10 pages across the dataset before you trust the verdict.

Trap 2: Forgetting that consolidation is a strategy, not a compromise. A good 301 redirect (301 是永久重定向) sends the link equity (链接权重) of the deleted pages to the surviving one. Done right, the surviving page ranks higher than either original page did. Done wrong — by redirecting to the homepage or to an unrelated page — you flush the equity down the drain.

Trap 3: Mistaking volume for value. A common reaction to a low audit score is "we need to write more." Most sites I've audited are suffering from too much content, not too little. The path to recovery is almost always deletion + consolidation, not new production.

The Frame That Stuck With the Client

At the end of the project, the CMO (Chief Marketing Officer, 首席营销官) asked me: "If you had to summarize the whole thing in one sentence, what would it be?"

I said: "Google rewards pages that earn their existence. Audit accordingly."

That framing has stayed with me because it captures something that's easy to miss when you're buried in scoring rubrics and spreadsheets. The pages that get demoted aren't necessarily bad in some absolute sense — they're bad relative to what the user could have gotten from any of the other results on the page. Every page on your site is competing with the entire web for attention. If yours doesn't bring something the others don't, you lose — not because Google is mean, but because the user has alternatives.

The audit works because it forces you to ask, page by page: does this one earn its spot? Most sites can't honestly answer "yes" to more than half their URLs. The ones that can, rank.