Citation Equity

The research behind each check

Short answer: none of the checks is proven yet. Treat the score as a writing checklist.

Each check below shows its grade, what the studies found, what our own test found (or, for seven checks, a second run with more pages), and links to the studies.

In our own test, applying every fix together raised the citation score by +0.049 (95% CI [+0.024, +0.076], n 78) over the original page and by +0.071 over a paraphrase of the same passages, possibly through length or query-term overlap.

The data is open: every result and score, with where each page came from and a hash of its text. The page text, edited versions, prompts and answers are not shared, since the pages belong to their publishers.

Paragraphs make sense alone

Untested

No study we opened tests whether paragraphs that make sense on their own get cited more.

Tested: no verdict by design

E1: descriptive, no verdict by design (d +0.007, 95% CI [-0.008, +0.023], n 24). E1 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Studies (2)

The check rests on how these systems work: many split a page into passages and pick passages one at a time, so a paragraph that opens with "This" or "It" can lose its meaning once lifted out.

What it checks. Paragraphs that still make sense when quoted on their own: they name their subject instead of opening with "This", "It" or "As mentioned", and run 40 to 150 words.

What the studies did. SAGEO Arena builds its search pipeline the way production systems are described: pages are cut into passages and passages are retrieved and reranked one by one. That explains why a self-contained paragraph could matter, but the benchmark does not test it. Zhang, He and Yao propose that useful pages have "paragraphs that each carry a distinct claim", as a hypothesis for later testing. A web search for a controlled test found only practitioner guides.

Weights: 1 standard · 0.15 by evidence

Headings ask questions

Mixed evidence

No controlled study we opened tests question-style headings, and the closest evidence does not favour them.

Tested: no clear effect · Not scored

E3: no effect larger than 0.03 (d -0.016, 90% CI [-0.028, -0.004], n 57, Holm p 0.164). E3 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Not scored: Headings ask questions. A test on 57 pages found no effect larger than 0.03, so it counts only if you give it a weight.

Studies (3)

Across three AI search platforms, cited pages in Q&A format shaped answers slightly less than other pages (about 6% lower), and a controlled study found that formatting-only changes made little difference to which page was cited.

What it checks. Section headings (H2 and H3) phrased as the questions readers ask, such as "How do I fix bitter espresso?".

What the studies did. Zhang, He and Yao compared cited pages in Q&A format with other cited pages and found a mean influence of 0.0947 against 0.1005, which they call "the important negative case": the question format is a wrapper, and what helps is the evidence inside. What Gets Cited ran matched pairs of pages that differed only in formatting (dense paragraphs versus organized sections) and found no consistent effect on the first citation. Martinez rates formatting alone as generalizing poorly and advises testing headings without assuming which way the effect goes.

A question heading can still help a reader scan the page. The evidence says not to expect the heading by itself to earn citations.

Weights: 0 standard · 0 by evidence

Lists, not long sentences

Mixed evidence

Pages that shape AI answers the most tend to have more lists, but a controlled test of formatting alone found no effect.

Tested: no clear effect · Not scored

E1: no effect larger than 0.03 (d -0.007, 90% CI [-0.027, +0.013], n 33). E1 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Not scored: Lists, not long sentences. A test on 33 pages found no effect larger than 0.03, so it counts only if you give it a weight.

Studies (3)

The list link may come from those pages being better overall, so use a list when the content really is a set of items or steps.

What it checks. Steps, options and other runs of three or more items written as a bulleted or numbered list instead of a long sentence.

What the studies did. In the observational study, pages in the top quarter by influence had a list density of 0.428 against 0.048 for the bottom quarter; they also had about twelve times as many headings and were about eleven times longer, so lists are one of many differences. What Gets Cited found no consistent effect from formatting-only edits. A preprint (GEO-SFE) restructured 200 articles, changing headings, chunking, lists, tables and emphasis together, and reported citation rate rising from 45.0% to 52.8% across six engines; lists were not tested on their own.

Weights: 0 standard · 0 by evidence

Real numbers

Mixed evidence

Real, checkable numbers help, but adding figures for their own sake helped in one early test and hurt in a larger one.

Tested: too few pages

E3: insufficient n (d -0.007, 95% CI [-0.045, +0.026], n 11). E3 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Studies (4)

Adding statistics raised visibility in the original GEO experiment, where the page was already one of five sources given to the model, but C-SEO Bench found that adding statistics lowered a page's citation rank in 19 of 24 settings. Pages that contain real numbers are used more in AI answers, so aim for relevant, checkable figures rather than more numbers.

What it checks. Concrete numbers with context, such as percentages, amounts with units, prices and counts. Bare years and list numbers do not count.

What the studies did. In GEO, an LLM rewrote one of the five sources to "include quantitative statistics instead of qualitative discussion, wherever possible", and that source's position-weighted share of the answer rose from 19.3 to 25.2. C-SEO Bench applied the same kind of rewrite across six domains and four engine models and measured citation rank instead: the paper reports that the rewrite lowered rankings in 19 of 24 evaluated settings (it does not say how those 24 divide across domains and models), and in the main experiment no rewrite helped in question answering. Observational data still links numbers to use: cited pages containing numbers had a 61.55% higher influence score. Martinez puts it plainly: the criterion is not to add numbers, but to give relevant, verifiable, dated and attributed evidence. A made-up figure may get reused and still mislead readers.

Weights: 1 standard · 0.25 by evidence

Comparison tables

Untested

No study we opened tests tables on their own.

Tested: too few pages

E3: insufficient n (d -0.027, 95% CI [-0.054, -0.001], n 11). E3 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Studies (3)

Comparison content, which tables often carry, was linked to more citations and more use in answers, while formatting-only changes showed little effect, so use a table to present a real comparison rather than to raise a score.

What it checks. Comparisons and lists of "Key: value" pairs laid out as a data table, so each value sits next to its label.

What the studies did. What Gets Cited found that a page without comparisons to alternatives lost the first citation more often (one of the factors significant in at least four of six models), but it tested the comparison content, not the table format, and formatting-only edits had no consistent effect. The observational study found that cited pages with comparison content had a 55.28% higher influence score; it reports no result for tables. Martinez advises testing tables without assuming which way the effect goes.

Weights: 1 standard · 0.15 by evidence

Quotes say who said them

Mixed evidence

Adding quotes helped in one early test, but a larger test found no gain, so quote people only where it adds something.

Tested: unclear, leaning positive

E3: inconclusive (d +0.026, 95% CI [+0.001, +0.053], n 62, Holm p 0.179). E3 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Studies (3)

Adding quotations gave the largest gain in the original GEO experiment, the source of the popular "up to 40%" figure, but C-SEO Bench found no significant gain from added quotes when it measured citation rank. That 40% is a bigger share of answer text in one fixed five-source setup, not more clicks or a better chance of being found.

What it checks. Quotations that name who said them, for example "says Jane Doe, head roaster at Example Coffee".

What the studies did. GEO's Quotation Addition rewrite raised the source's position-weighted share from 19.3 to 27.2, about 41% in relative terms; this is where "up to 40%" comes from. Martinez explains the limits: the page was already one of the five sources in the prompt, the metric is a share of answer text, and no clicks or traffic were measured; the survey rejects "GEO increases visibility by 40%" as a general claim. C-SEO Bench found no significant ranking gain from the Quotes rewrite in its main experiment.

A quote from a named expert can still add real information for readers. Attribute it properly, and do not expect the quote alone to lift citations.

Weights: 1 standard · 0.25 by evidence

Claims link a source

Mixed evidence

Linking credible sources helped in one early test, but later tests found little gain from adding references alone.

Tested: no clear effect · Not scored

E3: no effect larger than 0.03 (d -0.010, 90% CI [-0.023, +0.002], n 49, Holm p 0.228). E3 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Not scored: Claims link a source. A test on 49 pages found no effect larger than 0.03, so it counts only if you give it a weight.

Studies (4)

Adding citations to credible sources helped in the original GEO experiment, especially for lower-ranked pages, but later tests found little or no gain from adding references alone. Claims backed by concrete evidence did win citations in a controlled study, so link sources where they support a specific claim.

What it checks. Statistics and research claims backed by a link to their source in the same sentence or paragraph.

What the studies did. GEO's Cite Sources rewrite raised the share from 19.3 to 24.6, and when all five sources were rewritten the fifth-ranked page gained 115.1% while the top-ranked one lost 30.3%. C-SEO Bench found no significant ranking gain from its Citations rewrite in the main experiment. ConflictingQA edited pages on contested questions and found that adding scientific references had a neutral or negative effect on whether models sided with the page, while making the page more relevant to the question helped a lot. What Gets Cited found that claims backed by evidence such as tests or certifications beat the same claims without it.

Weights: 0 standard · 0 by evidence

Answer first

Mixed evidence

Some studies found that answering the question at the top helps, but the gains were narrow: an added top-of-page summary helped in only two product domains and made no difference for question answering or for one of the engine models.

Tested: no clear effect · Not scored

E3: no effect larger than 0.03 (d +0.012, 90% CI [+0.002, +0.022], n 66, Holm p 0.179). E3 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Not scored: Answer first. A test on 66 pages found no effect larger than 0.03, so it counts only if you give it a weight.

Studies (4)

All of the evidence is indirect (case studies, an added summary, rules learned from engines), so treat this check as a sensible habit rather than a proven lever.

What it checks. Each section opens with a short, direct answer of 8 to 40 words before the detail, without hedges like "It depends".

What the studies did. In SAGEO Arena's full pipeline, case studies found that putting the direct answer in early paragraphs improved reranking, while moving the answer later caused significant rank drops even though the answer was still on the page. In C-SEO Bench, a short summary added to the top of the page was one of only two rewrites with a significant gain, in the retail and video games domains; no rewrite made a significant difference in question answering, or for the Haiku 3.5 engine model. SAGEO Arena also found that in its shopping domain every rewrite it tested lowered citation, and that AutoGEO's rewrite (which applies "Conclusion First" among many other rules) caused the largest drop in being retrieved. In ConflictingQA, starting a page with a line saying which question it is about substantially raised how often models sided with it. AutoGEO's preference rules for Gemini, GPT and Claude all include "state the key conclusion at the beginning of the document". These test a page-level answer or summary; none tests an answer at the start of every section, which is what this check looks for.

Weights: 0 standard · 0 by evidence

Your questions answered

Mixed evidence

Relevance to the question is the most repeatable finding in the studies we opened: in controlled tests, edits that made a page address the question more directly made models use it more.

Tested: helped a little, maybe through wording overlap

E1: helps, possibly through length or query-term overlap (d +0.027, 95% CI [+0.010, +0.043], n 54, Holm p 0.006). E1 results

Replication: d +0.036, 95% CI [+0.016, +0.059], n 52.

Follow-up test (E2): the heading form adds nothing (a question heading against the same words as plain text: d +0.003 main, -0.001 replication, no effect larger than 0.03), and the question's words carry the gain (against a same-size heading without them: d +0.048 replication, helps; +0.017 main, inconclusive). Plain text with the question's words beat the untouched page by +0.023 (main) and +0.037 (replication), reported without a verdict.

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Studies (4)

This check measures a narrow stand-in for that (the question's key words in a short opening sentence), and in the largest controlled study, pages missing the question's terms lost the first citation in only four of six models. Answer the questions your readers really ask, in plain words, and do not repeat key words for their own sake.

What it checks. Each target question you enter gets a direct answer: the section whose heading matches it (or the paragraph that matches it best) opens with a plain sentence of 8 to 40 words that uses most of the question's key words. Applies only when you give target questions.

What the check looks for. You enter the questions your readers ask (the check does not apply without them). For each question, the check takes its key words (small words like "what", "is" and "the" are dropped, and endings like "-s" and "-ing" are trimmed), finds the H2 or H3 heading that shares at least half of them, or else the paragraph whose first sentence matches best, and counts the question as answered when that section opens with a plain sentence of 8 to 40 words that is not itself a question and contains at least 60% of the key words. These thresholds are ours.

What the studies did. ConflictingQA edited pages on contested questions one way at a time: edits that made a page more relevant to the question, such as adding a single sentence that makes its stance on the question obvious or starting it with "The following text is about the question", had a substantial positive effect on how often models sided with the page, while stylistic edits did not. What Gets Cited tested pages that lacked query terms the competing page had (its "Keyword Gap" factor): the page with the terms won the first citation, significantly in four of six models (Kimi K2, GPT-5 Nano, GPT-5 Mini, GPT-5.2) and not significantly in Gemini 2.5 Flash or Claude 3.5 Sonnet. In SAGEO Arena's reranking case studies, adding content that directly addresses the query's information need and placing the answer early raised a page's rank, and rewrites that mentioned the query's terms without addressing the question did well at retrieval but lost at reranking and generation. Martinez rates query-document relevance as strong in controlled settings, on the condition of a well-defined intent and no fabrication. None of these tests our exact rule, so the check is a writing habit with good reasons behind it, not a measured lever.

Weights: 0.5 standard · 0.25 by evidence

Compares the options

Mixed evidence

In a controlled study, product pages with no comparison to alternatives lost the first citation more often, but the effect was significant in only four of six models.

Tested: too few pages

E3: insufficient n (d +0.003, 95% CI [-0.020, +0.026], n 24). E3 results

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Studies (3)

An observational study found that pages with comparison content shaped AI answers more. No study tests comparison wording on its own, so compare options where readers really choose between them, with real differences, and skip it where they do not.

What it checks. Where readers choose between options, the draft compares them in so many words, such as "X vs Y", "brighter than", "unlike" or "instead of". Applies to drafts of 300 words or more.

What the check looks for. In drafts of 300 words or more, any heading, sentence or table row with explicit comparison language: "vs", "versus", "compared with" or "compared to", "than" (but not "more than 25 seconds"), "unlike", "instead of", "rather than", "alternatives", "pros and cons", "which is better", or a question heading of the form "X or Y?". One is enough to pass. The check never writes a comparison for you, because only you know the real differences.

What the studies did. What Gets Cited compared matched product pages that differed only in whether they compared the product with alternatives. The page with the comparison won the first citation with odds ratios of 7.45 (Gemini 2.5 Flash), 4.55 (Claude 3.5 Sonnet), 4.72 (Kimi K2) and 2.14 (GPT-5 Nano), all significant, and 1.78 (GPT-5 Mini) and 1.61 (GPT-5.2), which were not significant. The test used shopping-style pages with two sources per trial. Zhang, He and Yao found that cited pages with comparison content had a 55.28% higher mean influence score (0.1389 against 0.0894), and that comparison was the second most influential role a cited passage played in an answer; they list "adding comparison sections will increase future absorption" as a claim their observational data cannot establish. Martinez rates extractable evidence, including comparisons, as moderate to strong support when it is truthful and fits the reader's intent.

Weights: 0.5 standard · 0.25 by evidence

Your reader's words up top

Mixed evidence

Using your reader's own words where the page opens can help it get picked, but piling in query words is not the same thing.

Tested: helped a little, maybe through wording overlap

E1: helps, possibly through length or query-term overlap (d +0.035, 95% CI [+0.008, +0.066], n 33, Holm p 0.037). E1 results

Follow-up test (E2): with a different editor (Qwen3.8-27B) and judges (Gemma-4-12B and Qwen3.5-9B), the gain kept its sign (d +0.029, 95% CI [+0.003, +0.061], n 29).

Gemma-4-12B and Qwen3.8-27B judges, Gemma-edited.

Studies (4)

In an early controlled study (Aggarwal et al. 2024), a rewrite that added more of the query's keywords (Keyword Stuffing) lowered the page's share of the answer, 17.7 against 19.3 at baseline, and in the largest controlled study, pages missing the query's terms lost the first citation in only four of six models. Work the reader's words into the title or opening where they fit naturally, and leave them out where they do not.

What it checks. The key words of the page's questions appear in the title, the first heading or the first paragraph (a question section above them does not push them down); the score is the share found, averaged over questions. The questions are the ones you typed or imported, else the page's title and topic questions, never a section's. Common verbs such as "work" and "use" do not count. It names missing words as advice and never adds them for you.

What the check looks for. It measures the questions you typed or imported when there are any; otherwise the page's own questions, made from its title (when the title is a question) and its topic ("What is ...?", "How does ... work?"). It never uses a question made from a section heading, and without any question it does not apply. For each question it takes the same key words as the question-coverage check, leaving out common verbs ("work", "use", "take", "make"), and counts how many appear in the title (the first H1), the first heading after it or the first paragraph. A question section does not push those down: the check reads every heading down to the first one that is not a question, and every paragraph down to the first one outside a question section, so with the same questions, adding a question section never lowers the score. On a page whose sections are all questions, this reaches further down than the title, first heading and first paragraph that our test edited. A question asked twice counts once. The score is that share, averaged over the questions. The advice names up to three missing words per question and says where they could go; it never names a word that is already there and never writes the words in for you.

What the studies did. GEO's Keyword Stuffing rewrite ("Modifies content to include more keywords from the query, as expected in classical SEO optimization") lowered the source's position-weighted share of the answer from 19.3 at baseline to 17.7, and the authors find such methods offer "little to no improvement"; it added keywords anywhere in the page, not into the title and opening. What Gets Cited tested pages missing query terms that the competing page had: the page with the terms won the first citation, significantly in four of six models and not in Gemini 2.5 Flash or Claude 3.5 Sonnet. In SAGEO Arena, rewrites that mentioned the query's terms without addressing the question did well at retrieval but lost at reranking and generation. ConflictingQA found that opening a page with a line naming the question it answers substantially raised how often models sided with it. The rubric counts the stuffing result and the two null models as null or negative findings, so the grade is mixed.

Our own test. E1's candidate check for query wording worked the query's missing key terms into the title, the first heading and the first paragraph, adding only key terms and function words. Against a paraphrase of the same spans it raised the citation score by d +0.035 (95% CI [+0.008, +0.066], n 33, Holm p 0.037), possibly through query-term overlap; the replication set had too few valid edits for a verdict (n 12). In E2, with a different editor (Qwen3.8-27B) and different judges, the gain kept its sign (d +0.029, 95% CI [+0.003, +0.061], n 29). These are local models in a fixed context of four sources.

Weights: 0.5 standard · 0.25 by evidence

Your site checks

What the studies and the engines' own documentation say about each check on Your site. These are reported apart from the score.

Author name

Untested

No study we opened tests whether a byline changes how AI answer engines use or cite a page. Google asks publishers to make clear who created a page, with a byline where readers expect one, but says this is about trust and is not itself a ranking factor.

Sources (1)

Dates

Mixed evidence

A controlled study found that the date a passage states changes how language-model rerankers order it: newer dates moved up and older ones moved down. A date can therefore help a fresh page and hurt an old one, so keep the published and updated dates accurate rather than adding dates for their own sake.

Sources (1)

Structured data (schema)

Mixed evidence

Google says no special schema.org markup is needed to appear in its AI features, and a matched study of pages that added JSON-LD found no meaningful rise in AI citations. Structured data can still describe the page accurately for search engines; keep it matching the visible text. The detail also notes, for information, whether sameAs links the author or organization to their profiles elsewhere: Google documents sameAs for that, and no source we opened tests it for AI citation.

Sources (3)

Search can show it

Strong evidence

Google and Bing both document that indexing directives decide whether a page can appear in their AI answers: Google needs a page to be indexed and eligible for a snippet, so noindex, nosnippet and max-snippet:0 keep it out, and Bing keeps pages marked NOARCHIVE out of its chat answers.

Sources (2)

Main URL (canonical)

Untested

No study we opened tests whether a canonical URL changes AI citation. For search engines it tells which of several duplicate URLs to show and where to consolidate signals, which matters most when the same article lives at more than one address.

Sources (1)

AI crawlers allowed

Strong evidence

The answer engines' own documentation says a page their search crawler may not fetch is left out of their answers: OpenAI, Anthropic and Perplexity each name a search crawler, and Google asks that robots.txt allow crawling. Blocking a training crawler such as GPTBot or Google-Extended does not, by their accounts, remove a page from search answers.

Sources (4)

Text loads without JavaScript

Moderate evidence

A large measurement of crawler traffic found that the AI crawlers of OpenAI, Anthropic and Perplexity fetch pages without running JavaScript, so text that only appears after scripts run is invisible to them. Google renders JavaScript for Search and Gemini but still recommends server-side rendering because not all bots run it.

Sources (2)

Who gets cited

What the studies say about each kind of site AI answers cite, and the rules behind each Avoid line.

Forums and Q&A sites

Mixed evidence

Forums are heavily cited, but the studies disagree by engine and over time. Profound found Reddit the top source for Google AI Overviews and Perplexity; Ahrefs found Reddit and Quora high only in AI Overviews; Semrush saw ChatGPT's Reddit citations fall from close to 60% of responses to about 10% in September 2025. No study tests whether posting there gets your own page cited.

Sources (5)

Wikipedia and other wikis

Mixed evidence

Wikipedia is among the most cited domains in every study we opened, above all in ChatGPT, but Semrush saw ChatGPT's Wikipedia citations fall from about 55% of responses to under 20% in September 2025, and found it in about 2% of Google AI Mode responses. Whether you appear there is decided by Wikipedia's notability and conflict-of-interest rules, not by you.

Sources (5)

Video

Moderate evidence

YouTube is among the top sources for Google AI Overviews and Perplexity in Profound's and Ahrefs' data, and Semrush found it second in Google AI Mode. Ahrefs found it missing from ChatGPT's top ten. No study tests whether publishing a video gets your own page cited.

Sources (4)

Review platforms

Mixed evidence

The studies disagree. In Profound's data G2 is among the ten most cited sources of ChatGPT and Perplexity, and Yelp and Tripadvisor among Perplexity's, but Ahrefs' top tens for those two engines have no review site. No study tests whether more reviews get your own page cited.

Sources (4)

News and editorial sites

Moderate evidence

News and editorial sites fill much of ChatGPT's top ten in Profound's data (Forbes, Business Insider, New York Post, Reuters) and in Ahrefs' (Reuters, Apple News, AS.com). No study tests whether pitching gets your own page cited.

Sources (3)

Reference, health and government sites

Mixed evidence

The studies disagree. Ahrefs puts Mayo Clinic, Cleveland Clinic and the NIH in Google AI Overviews' top ten, and Semrush lists the NIH third in Perplexity, but Profound's top tens for those engines have no reference site, nor does Ahrefs' for Perplexity. These sites are cited for what they publish themselves.

Sources (3)

Professional and social networks

Moderate evidence

All three studies put a social network in Perplexity's top sources: LinkedIn in Profound's and Semrush's data, Instagram in Ahrefs'. LinkedIn is also among Google AI Overviews' top sources in Profound's data, and Semrush found it the most cited domain in Google AI Mode (nearly 15% of responses). No study tests whether posting there gets your own page cited.

Sources (5)

Shops and marketplaces

Mixed evidence

The studies disagree. Ahrefs puts Amazon in ChatGPT's top ten, but Profound's ChatGPT top ten has no marketplace, and no other top ten we opened has one. The review rules below apply to listings too.

Sources (3)

Blogs and other sites

Moderate evidence

Medium, a blogging platform, is in Google AI Overviews' top ten in Profound's data and third in ChatGPT in Semrush's. More to the point, the ten most cited domains together held only about 10% to 16% of each engine's citations in Profound's data, and no single domain more than 7.8%; the rest went to other sites, which is where your own page competes.

Sources (3)

Method and limits

A grade describes the research, not your draft. Scoring by evidence uses it as the weight: strong 1, moderate 0.7, mixed 0.25, untested 0.15. Not scored: Headings ask questions. A test on 57 pages found no effect larger than 0.03, so it counts only if you give it a weight. Not scored: Lists, not long sentences. A test on 33 pages found no effect larger than 0.03, so it counts only if you give it a weight. Not scored: Claims link a source. A test on 49 pages found no effect larger than 0.03, so it counts only if you give it a weight. Not scored: Answer first. A test on 66 pages found no effect larger than 0.03, so it counts only if you give it a weight.

Our own test applied each check's fix to real pages and asked two local models to answer from four sources. A second run asked again, with more pages, for the seven checks the first had too few pages for, and its result replaces the first for them. Neither changes a grade. d is the change in citation score, where a page cited first scores 1 and an uncited page 0. No effect larger than 0.03 means a small effect either way is not ruled out.

Strong evidence
Several controlled studies agree, including real search.
Moderate evidence
Studies point the same way; none found harm.
Mixed evidence
At least one study found no effect or harm.
Untested
No direct test yet.

Your questions and the searches behind each question are helpers for planning what to write. They are not part of the score, and we have not tested whether they change which pages AI cites.

Why the score isn't a forecast

  • Most tests hand the page straight to the AI; real search has to find it first.
  • Studies disagree: tricks that helped in one failed in larger ones.
  • Engines differ and change.
  • Our weights are our choice, and you can change them.

The full evidence document has the grading rules and the studies; the reading notes have the quotes.