There’s a kind of content that looks fine (H1, subheads, keywords, even FAQs)… but when a user lands, they don’t find anything new. And when a search engine evaluates it, neither does it.
That’s the real problem with thin content: it’s not “short text”, it’s low value. And today it’s not only Google filtering it out—AI answer engines also tend to ignore it (or simply won’t cite it).
What is thin content, and why does Google care?
Thin content is content that adds “little or no value” to the query intent: repeated templates, cookie-cutter pages, affiliate pages without added value, generic text, etc. Google’s spam policies explicitly mention thin affiliate pages: affiliate sites/pages that replicate content without providing additional value.
The modern way to think about it isn’t “penalty”—it’s eligibility. Ranking systems prioritize helpful, reliable, people-first content, not text created to manipulate rankings.
And with spam and low-quality-result updates, Google has strengthened its ability to take more targeted action against abusive practices that produce unoriginal or low-quality content.
What changes with AI in Google and in answer engines?
Two key ideas:
- There are no “special requirements” for Google AI Overviews / AI Mode: standard SEO best practices still apply, and there’s no mandatory “AI optimization” to appear there.
- In engines like ChatGPT Search, Perplexity, or Claude with browsing, content is selected to answer and (when relevant) cite. That favors pages with “citable-ready” units: crisp definitions, comparisons, verifiable lists, spec tables, dates, sources, etc.
If your page is thin, it may still be indexed… but it can lose:
- rankings (lower satisfaction / lower usefulness),
- coverage (better alternatives win),
- and citations (there aren’t reliable “blocks” to support an answer).
The Tacmind framework: Thin Content Radar (TCR)
To audit thin content with LLMs without relying on vibes, use a radar with five signals. I call it Thin Content Radar (TCR):
- Intent coverage: does it answer what the user came to solve?
- Information density: how many facts, steps, decisions, or criteria per paragraph?
- Differentiation: what’s here that isn’t a standard rewrite? (examples, data, method, expertise)
- Evidence & verifiability: precise definitions, sources, dates, boundaries, caveats
- Retrieval-ready structure: can engines cite it in chunks without extra context?
.png)
We tried a “quick detector” based only on word count.
It failed: it flagged short-but-perfect pages (glossaries) and let long-empty guides pass.
We fixed it with TCR: intent + density + evidence, at section level.
The result: fewer false positives and a much clearer fix backlog.
— Pablo López, Tacmind
How thin content detection by LLMs works in practice
You don’t need magic. You need a repeatable pipeline.
Step 1) Extract the main content (no nav/footer boilerplate)
If the LLM reads menus, CTAs, and repeated site text, your audit gets noisy. The goal is to evaluate the useful body.
Step 2) Split into evaluable chunks
AI systems work in fragments (chunks). Your audit should too. Practical rule: 150–300 words per chunk, or by H2/H3 sections.
Step 3) Define intent (don’t guess)
For each URL, store 1–3 intent statements as questions:
- “What is X and what is it for?”
- “How to choose X?”
- “X vs Y: which should I pick, and when?”
Step 4) Use the LLM as a judge (and design the judgment)
This aligns with research on using LLMs as evaluators (e.g., G-Eval and broader “LLM-as-a-judge” work). (arxiv.org)
Prompt template (per-section audit):
Role: SEO auditor + quality rater.
Task: Evaluate whether this excerpt provides unique value for the intent: "<INTENT>".
Return JSON with:
- intent_coverage (0-5)
- info_density (0-5)
- differentiation (0-5)
- evidence_verifiability (0-5)
- retrieval_readiness (0-5)
- why_thin (short list)
- fix_actions (concrete edits)
Content:
<PASTE SECTION>
Step 5) Add anti-hallucination controls
An LLM auditor can be wrong. Two simple safeguards:
- Two-pass scoring: evaluate with two prompts (or two models) and compare.
- Consistency checks: if the model “invented” facts, force it to score only what is explicitly in the excerpt. (Research on LLM consistency can inspire practical validation design.) (arxiv.org)
Scoring method: TTS-100 (Thinness & Trust Score)
TCR gives you signals. Now you need a number for prioritization.
- TTS-100 ranges from 0 to 100.
- 100 = excellent (not thin).
- 0 = extremely thin.
- Quick thresholds:
- <60: high risk (priority fixes)
- 60–79: improve to compete / be cited
- 80+: maintain and refresh
Prioritization: Impact × Thinness
With TTS calculated, prioritize with a simple formula:
Priority = (SEO impact + AI impact) × (100 − TTS) ÷ Effort
- SEO impact: organic traffic potential / associated revenue
- AI impact: likelihood of being cited (explainable topics, comparisons, “best X”, definitions)
- Effort: real hours (editing + expert review + design + QA)
We tried rewriting 50 URLs “big bang style” in parallel.
It failed: we spent weeks on low-impact pages and traffic didn’t move.
We fixed it with Impact × Thinness: quick wins first (high impact, very thin).
The result: less debating and more execution.
— Pablo López, Tacmind
A 30-day plan to eliminate thin content (without rebuilding the whole site)
Days 1–7: Audit + baseline
- Crawl + main-content extraction
- Chunking by H2/H3
- TCR + TTS-100 at scale
- Top 20 list: “high impact + low TTS”
- Baseline in Search Console + (if you track AI) baseline mentions/citations
If you want to measure citations and recommendations across multiple engines, use a hybrid tracking system (SERP + AI). Tacmind explains a dual-measurement approach here: SEO tracking: SERP metrics vs AI answer metrics.
Days 8–14: Quick wins (the biggest “thinners”)
- Remove filler intros (long openings without answers)
- Add definitions and criteria (lists)
- Turn “generalities” into concrete steps
- Add 1–3 credible sources or references for key sections
- Rewrite H2s into real user questions
Days 15–21: Consolidation (when thin is structural)
- Identify cannibalization: 5 URLs saying the same thing
- Merge into 1 pillar + redirects/canonicals as appropriate
- Re-score at section level to ensure the pillar crosses a healthy TTS threshold
(Note: Some issues go beyond “thin” into explicit spam-policy territory—e.g., publishing third-party content to exploit site reputation.) (developers.google.com)
Days 22–30: Citation-ready + governance
- Restructure for answer engines (citable blocks)
- Add “answer units”: definitions, pros/cons, comparisons, decision tables
- Editorial checklist: no article ships with TTS <80
- Review robots/AI crawling settings as part of your visibility strategy
To speed up rollout, you can start with a guided diagnosis (demo) and a prioritized action plan: Book a demo or see plans.
Common mistakes when using LLMs to detect thin content (and how to fix them)
- Mistake: “thin = low word count”
- Fix: score intent + density + evidence per section (TCR), not word count.
- Mistake: feeding the model the whole page with boilerplate
- Fix: extract main content; otherwise the LLM rewards repeated template text.
- Mistake: asking for a verdict without structure
- Fix: force JSON output with criteria, scores, and fix actions.
- Mistake: trusting a single pass
- Fix: two-pass scoring + human review for sensitive topics (especially YMYL).
- Mistake: fixing thin by “adding more paragraphs”
- Fix: add decisions: criteria, steps, boundaries, examples, and (where needed) sources.
We tried “fattening” articles with 600 extra generated words.
It failed: volume went up, usefulness didn’t; TTS barely improved.
We fixed it by adding evidence, comparisons, and decision criteria (not filler).
The result: more time on page and more “citable” sections for AI.
— Pablo López, Tacmind
Connecting thin content to LLM visibility (GEO/AEO)
If your goal is to be cited, beyond “not being thin” you need to be easy to retrieve and attribute:
- Google says standard SEO practices apply to AI experiences like AI Overviews/AI Mode.
- OpenAI documents its crawlers and user agents (e.g., GPTBot and OAI-SearchBot) so sites can manage access via robots.txt.
- Perplexity documents how to manage its crawlers via robots.txt.
- Anthropic explains its bots (e.g., ClaudeBot) and how site owners can block/allow them.
If you want to translate this into editorial and IA/architecture work:
- Content for LLMs: how to write for answer engines
- LLM SEO Blueprint
- Best Answer Engine Optimization (AEO)
- (For technical discoverability) llms.txt: what it is and how it fits
Quick pre-publish anti-thin checklist
- Does the answer appear within the first scroll?
- Does every H2 add a decision, criterion, or step (not filler)?
- Are there at least 3 “citable” blocks (definition, list, comparison, steps)?
- Is there evidence/verifiability where it matters (dates, sources, limits)?
- Could another page say the same thing by swapping the topic name? (if yes, you need differentiation)
- Estimated TTS-100 ≥80?
FAQ
Does Google use LLMs to detect thin content?
Google talks about “automated ranking systems” and spam/quality policies and updates, but you don’t need to know the exact model to act: build genuinely helpful content and quality signals. (developers.google.com)
How many words counts as thin content?
There’s no reliable threshold. Some short pages are excellent (glossaries) and some long pages are empty. Evaluate intent, density, and evidence (TCR).
Which LLM should I use as a “detector”?
Whichever is most consistent for your domain—but use safeguards: structured prompts, two-pass scoring, and human review on critical pages. The “LLM-as-a-judge” literature stresses bias mitigation and consistency improvements. (arxiv.org)
Does this work for ecommerce (category and product pages)?
Yes, but “value” looks different: buying guides, comparisons, real FAQs, complete specs, clear policies, and differentiators. In ecommerce, thin often comes from repeated templates.
What if I have thousands of programmatic URLs?
First, cluster by template and apply TTS by pattern. Then: consolidate, add unique value per cluster, or deindex what you can’t make useful (better than carrying “noise”).
How do I know I improved for AI (citations)?
Measure: mentions, citations, prompts where you appear, and which URL is cited. If you don’t have a system yet, start with a fixed prompt set and a weekly baseline, then automate.
Was this helpful?
