Why AI Won't Quote Your Page: Hedged Claims, Unbacked Superlatives and Generic Headings

We ran Rubric, the AI citability auditor we built, on our own site first. It flagged 103 of our 182 pages for unbacked superlatives. On the worst page, our generative engine optimisation guide, about 37 of the 48 flags were harmless: quoted search queries, URL slugs and format names. About nine were real claims with no source, and two had gone out of date.

So read the flags before you fix anything, and fix only the sentences that make a claim nobody can check.

What decides whether AI quotes your page?

Search rank comes first in AirOps' ChatGPT data, analysed by Kevin Indig. Pages at position 0 in ChatGPT's search results were cited 58% of the time, and pages at position 10 just 14% (Indig, 2026b).

In the same data, always-cited and never-cited pages "look nearly identical on most content metrics". So wording matters once you're in the running, and only in narrow ways. Our breakdown of what ChatGPT cites and why goes further on rank.

Does confident writing get cited more?

Dropping a heavy hedge helped in a 2026 lab test, but making text sound more confident didn't help in a 2024 one. Rahul Vishwakarma and two Sprinklr colleagues ran 252,000 two-source trials across six LLMs for a SIGIR '26 paper. A heavily hedged claim lost to the plain version in five of six models, and claims backed by evidence such as tests beat unsupported ones in four (Vishwakarma et al., 2026).

The test used whole anonymised product reviews and scored which source got cited first. Sprinklr sells marketing software.

At UC Berkeley, Alexander Wan, Eric Wallace and Dan Klein made web paragraphs sound more confident or added references. Both edits had "a neutral to negative effect" on how often models sided with the text (Wan, Wallace and Klein, 2024). They measured agreement, which isn't citation.

Do unbacked superlatives cost you citations?

No study we've found isolates them. In the Sprinklr paper, an "Overly Promotional" tone reached significance in only two of six models (Vishwakarma et al., 2026).

Lily Ray tracked 100 B2B "best software" queries in Google's AI Overviews at three dates between April and June 2026. In the 80 that showed an AI Overview, when a brand's own self-promotional listicle was cited, the brand was left out of the actual recommendation in 224 of 323 cases, or 69% (Ray, 2026).

Krishna Madhavan, Principal Product Manager at Microsoft Bing, writes that "Unanchored claims" leave "AI unsure how to classify it". His fix is to "anchor claims in measurable facts": "42 dB dishwasher designed for open-concept kitchens" instead of "quiet dishwasher" (Madhavan, 2025). Google's helpful-content checklist asks whether a page title avoids "exaggerating or being shocking in nature" (Google Search Central, 2026b).

Does a heading like "Overview" cost you anything?

You're giving up the heading's chance to match the question. Kevin Indig calls heading-to-query match "the strongest content signal" in AirOps' 815,484-row ChatGPT dataset. Pages whose heading matched the query at 0.90 similarity or above were cited 41% of the time, against 30% below 0.50 (Indig, 2026b).

The finding is correlational, and AirOps sells AI content software.

Microsoft suggests swapping "a vague heading like 'Learn More'" for a question such as "What Makes This Dishwasher Quieter Than Most Models?" (Madhavan, 2025). A question mark on its own won't do it. The heading has to match what people ask, with a real answer under it. That's the core of answer engine optimisation.

What does Rubric's Content citability signals card count?

It counts pages. Each chip shows how many pages have at least one flag, and a red chip means at least one page is flagged. The card is advisory and doesn't feed your score. You'll find it under "Show more analyses" on your report's Overview.

Rubric estimates how quotable a page is. It doesn't count citations. Elisha Bajemon and Andre-Louis Rochet, who've built a content score themselves, call scores like this "a quality filter rather than a citation predictor" (Bajemon and Rochet, 2026).

Pair the card with Quotable passages, which lists your least quotable text blocks with one fix each. You can run the crawl from a free account.

Our Content citability signals card from the 5 October 2026 crawl: the eight signal chips and the top three pages in its flagged list.

What does each chip mean?

Each chip names a kind of text or markup that makes a page harder to quote or check.

Chip	What it flags	Our pages
Hedged claims	Claims softened with words such as "may" or "up to"	0
Unbacked superlatives	Words such as "best" or "the only" with no source in the same sentence	103
Generic headings	Stock headings such as "Overview" or "FAQ"	95
Unresolved passages	Passages that open on "it" or "this" without naming what they're about	3
Over-broad articles	Long articles with 12 or more subheadings and 2,000 or more words	89
Stale or mismatched schema	Structured data that's badly formatted or different from the page	10
Price hidden until JS	A price that only appears once JavaScript runs	0
Oversized passages (exp.)	Very long passages, a check Rubric marks experimental	3

What did Rubric find on our worst page?

Rubric's check flagged a search query we'd quoted, "best Shopify CRO agencies UK", and slugs such as /blog/best-ab-testing-tools-2026, though neither needs a source. The real claims read like "Profound is the best paid layer", with no source in the sentence. Two of them called Bing Webmaster Tools "the only first-party surface" without a source, and they're both out of date now.

A pattern match can't tell a quoted query from a brag, so someone has to read the list.

Use Claude and the Rubric MCP to sort every flag into keep, source or cut

Connect Rubric's hosted MCP to Claude, pull your audit with get_audit and open the worst page with get_page. Claude can sort the sentences behind the flags and draft rewrites for a person to approve. Then check_draft tests the rewritten page before it goes live.

Create a key on Rubric's MCP page and copy it straight away, because it's shown once. The same page gives you a one-line Claude Code command with your key filled in. Reads are free. start_crawl runs a live crawl of your site, and check_draft runs Rubric's checks on the text you paste in.

Paste this prompt:

Use list_audits to find my latest Rubric audit, then get_audit.
Run get_page on the page with the most unbacked superlatives.
Sort each flagged sentence into three groups:
1. Not a claim (a quoted query, URL or title). Keep it.
2. A claim we can back. Rewrite it with the figure or a named
   source in the same sentence.
3. A claim we can't back. Suggest a cut.
Show each rewrite beside the original. Publish nothing.
When I approve, run check_draft on the full page text.

Check every figure Claude adds against its source, because a fluent rewrite can carry a number nobody's measured. /mcp__rubric__help lists the tools, and /mcp__rubric__fix runs Rubric's side of the loop.

Put the figure and its source in the sentence

Replace each hedge with the figure it stood in for, and give each superlative a named source in the same sentence or cut it. Generic headings become the question the section answers.

Flag	Fix	Evidence
Unresolved passages	Name the subject in the first sentence instead of opening on "it" or "this"	Adding context to each chunk cut top-20 retrieval failures by 35%, or 67% with BM25 and reranking, in Anthropic's lab tests (Anthropic, 2024). Bing: "Key statements do not rely on implied content" (Microsoft Bing, n.d.)
Over-broad articles	Split an article that answers several separate questions, and link the parts	Bing: "URLs focused on a primary topic are more likely to be selected for grounding results" (Microsoft Bing, n.d.)
Stale or mismatched schema	ISO 8601 dates (Google, 2026), a plain numeric price with priceCurrency (schema.org), and values that match the page	Google lists "Making sure your structured data matches the visible text on the page" (Google Search Central, 2025)
Price hidden until JS	Put the price in the HTML the server sends	OpenAI's, Anthropic's and Perplexity's crawlers didn't render JavaScript in Vercel's December 2024 analysis (Vercel, 2024)

Stop splitting before you reach a page per query. Our long-tail AEO guide shows how to cover follow-up questions without one.

How do you rewrite a hedge or an unbacked superlative?

If there's no figure behind a hedge, the claim isn't ready. For superlatives, Rubric reads attribution literally: "according to Ahrefs" in the sentence clears the flag, while a trailing "(Ahrefs, 2026)" doesn't.

Illustration: a hedge. The before line is made up for this example.

Before: "Faster pages can often lead to big gains in revenue."

After: "Our page-speed fixes cut Bee Friendly Skincare's load time by 2.24 seconds, and its revenue went from $48,000 to $1,447,225 a year."

Paraphrased illustration: a superlative on our GEO page. Only the quoted words in the before line are verbatim.

Before: Bing Webmaster Tools' AI Performance report is "the only first-party surface" for AI citation data.

After (our proposed rewrite): "Bing Webmaster Tools' AI Performance report shows which pages Copilot cites. Microsoft Clarity's Citations report has been generally available since 13 May 2026 (Rizk, 2026), and Google says its Search Console generative AI report reached all websites worldwide on 31 August 2026 (Maoz and Samet, 2026). Google's report counts impressions only."

No word list catches every brag, so read your own pages as well. In sales copy, put the figure where your copywriting framework asks for proof.

What should a generic heading become?

A generic heading should become the question the section answers, in the words a reader would type. Kevin Indig's example is "When did SEO start?" over "The History of SEO", because "the very first word of the answer is SEO" (Indig, 2026a). To illustrate, "Introduction" becomes "Why won't ChatGPT quote my product page?", and "FAQ" becomes "What do buyers ask about delivery to Ireland?"

Rubric's list includes "FAQ" and "Frequently asked questions", yet Bing recommends "FAQ sections" in its AI Performance guidance (Madhavan et al., 2026). Keep the section and rename the heading.

If the heading's part of a template, fix it there:

• Webflow: rename the heading on the CMS template, or bind it to a field so each item can carry its own question.
• Shopify: a theme-section heading is shared by every product on that template, so put per-product questions in a metafield.
• WordPress: fix the block pattern or page-builder template, then check the posts already built from it.

Re-crawl, reopen the card, then check real answers

Re-crawl with start_crawl, wait for crawl_status to finish, then run compare_audits to see what moved. Reopen the Content citability signals card too and compare its page counts with the last crawl, because the card isn't part of the score.

After that, ask ChatGPT and Perplexity the questions that matter to you, and log which domains each answer cites. Our guide to AI visibility tracking covers tools that automate those runs. For Copilot, Bing Webmaster Tools shows first-party numbers Rubric can't see.

In Bing's AI Performance sample for the 90 days to 25 September 2026, GoGoChimp was cited 3,919 times for the grounding query "optimize content for AI retrieval pipelines", a 10.04% share of the citations Bing showed for it. The report covers Microsoft Copilot, Bing's AI summaries and partner tools.

Run a Rubric audit, open the card and read your worst page's flags before you change a word. Our AI SEO strategy for 2026 covers the rest.

Questions about hedges, superlatives and generic headings

Does the Content citability signals card change my Rubric score?

No. The card is advisory and sits outside the score, which was 89 on our 5 October 2026 crawl. Its flags show where pages are hard to quote or check, and any effect on citations shows up only in real answers. Track those in Bing Webmaster Tools and by hand.

Should I delete every superlative Rubric flags?

No. On our worst page, about 37 of 48 flags were quoted search queries, URL slugs and format names, and those can stay. A real claim needs a named source in the same sentence or a figure that proves it. If you can't find either, cut the claim.

Why does Rubric flag a sentence that only quotes a search query?

Because the check reads words and can't see what a sentence means. It looks for terms like "best" or "the only", and clears a sentence only when attribution such as "according to" sits inside it. Treat the list like a spellchecker's: it finds words, and you decide which are wrong.

Does adding statistics and sources get a page cited more by AI?

Not on current evidence. The 2023 GEO benchmark reported 30-40% visibility gains for its top three methods, which included adding statistics and citing sources (Aggarwal et al., 2024). A September 2026 preprint found "none of the three strongest 2023 GEO levers moves citation" across ten engine families (Bajemon and Rochet, 2026). Add evidence so readers can check it.

Does fixing stale schema get you cited by AI?

Fixing stale schema is accuracy work. Ahrefs compared 1,885 pages that added JSON-LD with 4,000 controls. Google AI Overview citations fell 4.6%, while AI Mode's 2.4% and ChatGPT's 2.2% rises were both "statistically indistinguishable from zero" (Linehan, 2026). Fix dates and prices so the schema matches the page.

Do original data and tables get a page cited by AI?

No study we've found shows they cause AI citations. Google says unique, useful content will "likely influence" your presence in generative AI search "in the long run" (Google Search Central, 2026a). That's guidance rather than measurement. Cyrus Shepard's study of 400+ sites linked proprietary assets to Google organic traffic gains and found "Uniqueness of information" had "no bearing on the data" (Shepard, 2026). It measured Google traffic, and AI citations weren't part of it.

References

• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024, v3 28 June). GEO: Generative Engine Optimization. KDD 2024 (the 2023 GEO benchmark). https://arxiv.org/abs/2311.09735
• Anthropic. (2024, 19 September). Introducing Contextual Retrieval. https://www.anthropic.com/engineering/contextual-retrieval
• Bajemon, E., & Rochet, A.-L. (2026, 7 September). Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End. arXiv preprint re-testing the 2023 GEO levers on ten engine families, with content-only scores repositioned as quality filters. https://arxiv.org/abs/2609.07559
• GoGoChimp. (2026, 5 October). Rubric report for gogochimp.com, hosted crawl of 182 pages: Content citability signals. First-party.
• GoGoChimp. (2026, 28 September). Bing Webmaster Tools AI Performance AI search queries export for gogochimp.com, data through 25 September 2026. First-party, unpublished.
• Google. (2025, updated 10 December). AI features and your website (Google Search Central). https://developers.google.com/search/docs/appearance/ai-features
• Google. (2026, updated 8 September). Article structured data (Google Search Central). https://developers.google.com/search/docs/appearance/structured-data/article
• Google. (2026a, updated 10 July). Optimizing your website for generative AI features on Google Search (Google Search Central). https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
• Google. (2026b, updated 5 October). Creating helpful, reliable, people-first content (Google Search Central). https://developers.google.com/search/docs/fundamentals/creating-helpful-content
• Indig, K. (2026a, 17 February). The Science Of How AI Pays Attention. Search Engine Journal. https://www.searchenginejournal.com/the-science-of-how-ai-pays-attention/567597/
• Indig, K. (2026b, 14 April). Shorter, Focused Content Wins In ChatGPT (data: AirOps, 815,484 query-page rows). Search Engine Journal. https://www.searchenginejournal.com/shorter-focused-content-wins-in-chatgpt/571857/
• Linehan, L. (2026, 11 May). Ahrefs study of schema and AI citations (1,885 pages, 4,000 controls). https://ahrefs.com/blog/schema-ai-citations/
• Madhavan, K. (2025, 8 October). Microsoft Advertising blog post on optimising content for inclusion in AI search answers. https://about.ads.microsoft.com/en/blog/post/october-2025/optimizing-your-content-for-inclusion-in-ai-search-answers
• Madhavan, K., Merchant, M., Canel, F., & Nigam, S. (2026, 10 February). Introducing AI Performance in Bing Webmaster Tools Public Preview. Bing Webmaster Blog. https://blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview/
• Maoz, H., & Samet, M. (2026, 3 June, rollout note dated 31 August). Google Search Central blog post on the Search generative AI performance reports. https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports
• Microsoft Bing. (n.d.). Bing Webmaster Guidelines (rewrite reported 27 February 2026). https://www.bing.com/webmasters/help/webmaster-guidelines-30fba23a
• Ray, L. (2026, 17 June). Substack study of self-promotional listicles in Google AI Overviews (100 B2B queries, 80 with an AI Overview). https://lilyraynyc.substack.com/p/why-calling-yourself-the-best-could
• Rizk, I. (2026, 13 May). Microsoft Clarity blog post announcing that Citations is generally available. https://clarity.microsoft.com/blog/citations-now-generally-available/
• schema.org. (n.d.). price property. https://schema.org/price
• Shepard, C. (2026, 9 April). 5 Data-Backed Features Of Websites Winning Google in 2026. Zyppy. https://signal.zyppy.com/p/winning-google
• Vercel, with MERJ. (2024, 17 December). Vercel analysis of AI crawler behaviour and JavaScript rendering. https://vercel.com/blog/the-rise-of-the-ai-crawler
• Vishwakarma, R., Kumar, S., & Jamidar, R. (2026, 25 May). What Gets Cited: Competitive GEO in AI Answer Engines. Proceedings of SIGIR '26. https://arxiv.org/abs/2605.25517
• Wan, A., Wallace, E., & Klein, D. (2024, v2 9 August). ACL 2024 paper on which evidence language models find convincing (ConflictingQA). https://arxiv.org/abs/2402.11782