{
  "license": "CC BY 4.0",
  "placement": "Insert the two paragraphs before the existing What this means for you explanation. Retain all original text and the existing first-paragraph direct answer. Apply identical modification-date convention to both modified arms.",
  "edits": [
    {
      "title": "Which AI visibility tools can a buyer verify?",
      "block": "measurement-1",
      "arm": "treated",
      "addition": "Check the event behind a visibility number before treating it as evidence about a source. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) found qualifying own-domain citations in 2,608 of 8,085 named vendor-answer pairs, or 32.3%, in the September 2026 B2B software archive. The base is named pairs, rather than every attempted answer or every citation.\n\nThe study states, “A citation describes a source choice.” A third-party source can accompany a vendor name without qualifying as that vendor's own domain. Ask whether the tool preserves that distinction and the underlying observation. This descriptive result does not compare monitoring products or establish that any particular tool is accurate. It illustrates why an export's event definition belongs beside its displayed rate.",
      "addition_words": 118,
      "before_sha256": "2775308c1321ecd3fa17f6cdd0362f14430f5259fb3ca3f5cc0dbd889dc4febf",
      "url": "https://docs.broadcastwell.com/answers/measurement/which-ai-visibility-tools-can-a-buyer-verify/",
      "after_text_sha256_lf": "d6ebf0a9d57bedbfdf141dac9dfd08ac5be717ba7223a7570379982c21ce20a0",
      "addition_sha256_utf8_lf": "c4a3852c76cff94265bb4e726722b1ab2d8346beed14463984912595e77c744b"
    },
    {
      "title": "What can Search Console show about AI visibility queries?",
      "block": "measurement-1",
      "arm": "placebo",
      "addition": "Use the query table for the question it can answer: which reported searches suggest a topic for investigation. Keep the reported wording, dates and metrics together so that a reader can identify the demand signal being discussed. If the question concerns an engine's answer, preserve that observation separately rather than substituting a search-table entry.\n\nApply the same care when describing a missing event. A filtered query-table result should retain its own scope instead of becoming a statement about the whole site's traffic. Where tables use different populations or aggregation rules, keep that difference explicit. Share only an aggregate that preserves the accounting when a raw query contains personal or commercial detail; the person's identity is not needed to explain the reported table.",
      "addition_words": 122,
      "before_sha256": "c925ec34a5b2f311bc42bb1767d7e62959c9767ff5708fca10c294a2fb19e1a0",
      "url": "https://docs.broadcastwell.com/answers/measurement/what-can-search-console-show-about-ai-visibility-queries/",
      "after_text_sha256_lf": "df4c1bc97eaa603a3edecfe00dbbd57b7514f40352e94708c847f45a7f2404a2",
      "addition_sha256_utf8_lf": "51e192c46aa61ea691f11b4af8ba1799f04365d3612003e607c74ee4b040ad99"
    },
    {
      "title": "What should a client AI visibility report contain?",
      "block": "measurement-2",
      "arm": "treated",
      "addition": "Keep an answer-level base beside a citation-level base in the report. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) retained 2,095 eligible target answers from the September 2026 B2B software archive; 39 had empty citation arrays, or 1.9%. Those answers contribute to the answer denominator but supply no entries to a citation-occurrence denominator.\n\nThe study states, “A citation describes a source choice.” A source count therefore needs its own definition, alongside the answer count and the exclusions. These observations describe the archived questions across five engines. They do not measure client visits, sales or the effect of content work. Preserve those boundaries when writing a result sentence so that an empty source array does not disappear merely because another table counts citations.",
      "addition_words": 120,
      "before_sha256": "c02460b845fd63e678beadbb31e2e0d74d95fa18b11a48262267f37cd6963231",
      "url": "https://docs.broadcastwell.com/answers/measurement/what-should-a-client-ai-visibility-report-contain/",
      "after_text_sha256_lf": "2263f2a0e245a82da7682a81b7dd30c444755a9e30d2277467be99cf87a6aa5e",
      "addition_sha256_utf8_lf": "d17ebde8e6107ea0aadb205eaa3d38ffcd09582663c934808ddb1c3b890d66dc"
    },
    {
      "title": "How can I check a tool's historical AI visibility data?",
      "block": "measurement-2",
      "arm": "placebo",
      "addition": "Read each historical point as a dated record with a definition. Establish whether its underlying answer was collected at that time, calculated later from retained answers or reconstructed under a changed method. Preserve those distinctions when the earlier and later periods appear beside one another. A line on a chart does not resolve an unavailable underlying record.\n\nKeep the original file when a calculation changes, and describe the new calculation without silently changing what the earlier report meant. Check whether the same buyer questions and engine set support the comparison. If they do not, present separate dated series or state the unresolved boundary. Ask which records remain available for inspection after access ends so that a later reader can understand the historical point.",
      "addition_words": 123,
      "before_sha256": "af27366fe211d0b0239929fcfa67b67eb8fad7fd9cd19006167c99dd6eeb9568",
      "url": "https://docs.broadcastwell.com/answers/measurement/how-can-i-check-a-tool-s-historical-ai-visibility-data/",
      "after_text_sha256_lf": "88b832f0a60bcf95f438d4e15613c0a594fb11bef40373ed1ccaf39207078644",
      "addition_sha256_utf8_lf": "edde0e35f7437c2c30293fcc15144979906741fc41cb1dff5060f389860d20f3"
    },
    {
      "title": "How should I check an AI baseline before changing content?",
      "block": "measurement-3",
      "arm": "treated",
      "addition": "Fix the citation rule as well as the question list before changing content. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) found 39 empty citation arrays among 2,095 eligible target answers, or 1.9%, in the September 2026 B2B software archive. A rule that silently drops those answers changes the base before any later comparison begins.\n\nThe study states, “A citation describes a source choice.” Keep source selection separate from support for a sentence and from a textual brand mention. These descriptive counts do not establish baseline stability or the required number of repeats; the repeat evidence above addresses a different question. They show a concrete denominator decision to freeze and retain when the same questions are measured again after a page change.",
      "addition_words": 120,
      "before_sha256": "20f5415572dd88e390d09303995d1553b7d00d6578cf88d1e662e6213ad7110b",
      "url": "https://docs.broadcastwell.com/answers/measurement/how-should-i-check-an-ai-baseline-before-changing-content/",
      "after_text_sha256_lf": "672521149c19e4402ad5c6549606038dfe71f7a256e0f8a690436ac900ff00d6",
      "addition_sha256_utf8_lf": "b1b8af3c6ceb157fcbb6db34a6549114e83ae75c5d0a6ae1a92ba6d172db7e2d"
    },
    {
      "title": "How can a buyer verify the conclusion?",
      "block": "measurement-3",
      "arm": "placebo",
      "addition": "Read the conclusion together with the dates and the retained classification. The answer quotation and the later source check describe different observations. Preserve the distinction between an incorrect historical statement and one classified as stale rather than presenting both as a single current-price conclusion. A later review date does not change the capture date of the answer.\n\nKeep each flagged statement identifiable when several observations support it. A question, a repeat and a supporting quotation remain evidence for the statement; they need not become additional flagged statements. The reference field should lead to the page relevant to that particular claim. Where a statement remains unverifiable, retain that label beside the other outcomes instead of converting uncertainty into either a confirmed error or a zero.",
      "addition_words": 124,
      "before_sha256": "67745ee8279b46e942a1487e7a72cbd97e328f61aaf3756c279b2a59b22cb3b4",
      "url": "https://docs.broadcastwell.com/answers/measurement/how-can-a-buyer-verify-the-conclusion/",
      "after_text_sha256_lf": "20d11b0e3d17c7c7696f4a22836c4df891bd1671daf93ca4f4bd9cc468d055bc",
      "addition_sha256_utf8_lf": "e6f35c66f9d33e198a25fee8d1f61fc35416a9681f251569bc76320ca0fcb4ec"
    },
    {
      "title": "How do we compare citation metrics when changing AI tools?",
      "block": "measurement-4",
      "arm": "placebo",
      "addition": "Start the comparison by matching the event and the base, then check how the number is stored. A naming measure and a cited-domain measure describe different observations even when a display uses the same general label. Preserve the definitions before joining the earlier and later reports so that the comparison does not depend on a familiar column name.\n\nKeep an unmatched field unresolved when the necessary definition or supporting records are unavailable. A prominence measure cannot be reconstructed merely by renaming another column. The collection conditions also belong beside the comparison: questions, engines, dates, repeats and exclusions define what the record covers. If those conditions change, describe the new series without presenting the definitional change itself as a client's observed gain or loss.",
      "addition_words": 123,
      "before_sha256": "07bd12f847d6d941466d0688cc36d49133518f48cc5edda1d1c5b0d61a348c52",
      "url": "https://docs.broadcastwell.com/answers/measurement/how-do-we-compare-citation-metrics-when-changing-ai-tools/",
      "after_text_sha256_lf": "008635e985d28fc888863978c70bdb6dd6d74b90ebbb0ed8b090a99679b06e41",
      "addition_sha256_utf8_lf": "f1acdd4cc7a362f9167480b858842986eb1428eb817881af4d1674f96a0b750f"
    },
    {
      "title": "How many repeated runs make an AI visibility rate useful?",
      "block": "measurement-4",
      "arm": "treated",
      "addition": "Retain a no-citation answer when calculating a citation rate across repeats. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) recorded 39 empty citation arrays among 2,095 eligible target answers, or 1.9%, in the September 2026 B2B software archive. Removing these answers would make the denominator depend on whether a source was returned.\n\nThe study states, “A citation describes a source choice.” The archived count is not an estimate of repeat agreement and supplies no universal repeat threshold. Its relevance is the outcome rule: repeated observations need the same eligibility definition, including valid returns with no source. Keep the repeat count, observation dates and exclusions beside any later rate. A different collection can have a different empty-array share without demonstrating a content effect.",
      "addition_words": 120,
      "before_sha256": "bcaf4372173e5111a26c9509a86a1fbfa90a4590f5211e51904d2293fac5fd3a",
      "url": "https://docs.broadcastwell.com/answers/measurement/how-many-repeated-runs-make-an-ai-visibility-rate-useful/",
      "after_text_sha256_lf": "048efb342ada32fe2f142527d43901abae9f6adf170d41e6f1467842ac777b72",
      "addition_sha256_utf8_lf": "4629b0ded0ab29013d72b5dd036b6acb4a0697279fedd537bdffa7ea506aae99"
    },
    {
      "title": "How does an AEO or GEO audit differ from an SEO audit?",
      "block": "measurement-5",
      "arm": "placebo",
      "addition": "Match the audit task to the observation being discussed. If the question concerns a buyer criterion in an answer, preserve the passage that states it and where it appears in the response. A record of whether a supplier name appears would leave that criterion unexamined. The relevant evidence is the captured wording rather than a broader audit label.\n\nKeep passages and answers as separate counting units. Several retained quotations can come from a single returned answer, so they should not silently become additional answers. The technical page checks address a different part of the work: whether a page can supply accessible, supported information. Keep those checks beside the answer evidence while preserving the distinction between the page's technical state and the buyer criterion actually observed.",
      "addition_words": 125,
      "before_sha256": "ba1ac8419209dd6a33b0e07649184b6d602b8dd5304943e880c5ccd28e00a241",
      "url": "https://docs.broadcastwell.com/answers/measurement/how-does-an-aeo-or-geo-audit-differ-from-an-seo-audit/",
      "after_text_sha256_lf": "9df064ac841a92cc1880258848ac6705fe96f238765e6d936212c490112a68c4",
      "addition_sha256_utf8_lf": "123f993a31e4ec8a2ccf8953b53245d920b4282503a3bd8bf20a5280049b4682"
    },
    {
      "title": "Can AI answers get company facts wrong?",
      "block": "measurement-5",
      "arm": "treated",
      "addition": "A cited source still needs to be checked against the particular company claim. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) counted qualifying own-domain citations in 2,608 of 8,085 named vendor-answer pairs, or 32.3%, in the September 2026 B2B software archive. That classification records naming and domain matching, rather than whether a feature sentence is true.\n\nThe study states, “A citation describes a source choice.” It follows that the source-choice count cannot certify the description attached to a name. Inspect the dated passage and preserve any unresolved mismatch. These data do not estimate a company-fact error rate or show that own-domain citations are more accurate than third-party references. The point is to keep the observed link and the factual verification as separate records.",
      "addition_words": 122,
      "before_sha256": "290dfc95ea255afceb67356f97fa010b10a4cd652fdf74c52df328534b1afb8e",
      "url": "https://docs.broadcastwell.com/answers/measurement/are-you-guys-starting-to-see-ai-backfiring-on-companies/",
      "after_text_sha256_lf": "5f49615bfab70245a6bcc71cf421e1cdfc62f7f36dfbb588b4906e508eafc84b",
      "addition_sha256_utf8_lf": "ee241874ed9fdedc4b3978e06dd92e8d6d9a48e54c3afd5f9a469e487d9f49b9"
    },
    {
      "title": "Can I use AI to do my SEO?",
      "block": "geo-basics-1",
      "arm": "placebo",
      "addition": "Verify the input claim before using a generated draft as the basis for an editorial decision. A mistaken allowance can change a comparison before any decision about its title, headings or links. Follow the cited reference and check the precise proposition rather than treating the presence of a citation as a completed factual review.\n\nKeep the answer capture date separate from the later product-page check. A historical reference can explain why a study classified a statement without establishing today's terms. Preserve the distinction between a numerical ceiling and an unlimited allowance when that is the disagreement under review. A broad description of a generous plan would lose the detail needed to understand the source comparison. The next editorial step follows the checked claim and its scope.",
      "addition_words": 126,
      "before_sha256": "8e906db7995fecd4db7f6f30003abb5486a223da421558767698456631c7c011",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/can-i-use-ai-to-do-my-seo/",
      "after_text_sha256_lf": "a8b9c13574a658cab080fe5b33114f5655ab8d7ffd88f2cd4bcf06f142d6b64b",
      "addition_sha256_utf8_lf": "9e7f958996a377797f05481e7192c24db81ab87c8c9358e103930995f4114eb8"
    },
    {
      "title": "Do brand-name prompts measure discovery?",
      "block": "geo-basics-1",
      "arm": "treated",
      "addition": "Keep the question condition attached when reporting whether a vendor was named or cited. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) found qualifying own-domain citations in 2,608 of 8,085 named vendor-answer pairs, or 32.3%, in the September 2026 B2B software archive. The base is already-named pairs under the archived matching rules.\n\nThe study states, “A citation describes a source choice.” That conditional percentage cannot be read as a generic discovery rate, and supplying a brand in a question changes the condition being investigated. Preserve the question text beside the result. The study is descriptive and does not test the causal effect of adding a brand to a request. It provides a concrete example of why a conditioned denominator needs a conditioned interpretation.",
      "addition_words": 122,
      "before_sha256": "07b4670c11b5c2d27dc5620f062846097654a841eb3fac7ad3ec4a56d6eea12a",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/do-brand-name-prompts-measure-discovery/",
      "after_text_sha256_lf": "ecf0d2ebbb940b687f103ed63a8af56f36c231fb442c747bbf1e906d3206d0e1",
      "addition_sha256_utf8_lf": "c1c656b34289b1cfa5c0e86b30bd16ac86d6f8d5c47f7ad439c65b3763a22c01"
    },
    {
      "title": "Does AI search replace SEO?",
      "block": "geo-basics-2",
      "arm": "treated",
      "addition": "Separate visible page characteristics from evidence that an optimization caused selection. Our [registered page study](https://docs.broadcastwell.com/research/studies/optimized-pages-in-ai-answers/) classified 3,913 of 4,420 successfully captured target URLs as definitely optimized, or 88.5%, using its stated precedence rules in October 2026. The URLs came from citations in the September B2B software archive.\n\nThe study states, “Presence does not prove consumption.” Its categories describe observed signatures on later-captured pages; they are not treatment assignments or a comparison with pages that were never cited. The result therefore cannot establish that AI search replaces SEO or that adding a signature increases citation probability. Keep technical access, factual page quality and the later answer observation as distinct checks. This source supplies descriptive context for that separation, rather than a universal publishing prescription.",
      "addition_words": 126,
      "before_sha256": "28d37990b0e5e37f62051f060eddd28032cff3428bb8ec83e18976f50ac6360b",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/does-ai-search-replace-seo/",
      "after_text_sha256_lf": "b4558ead8320e49d4f9ffd14e6c974d3d4161aa355eac63a01e1b140d7fa9f53",
      "addition_sha256_utf8_lf": "77f895d417897bf21104c1f6d2e2c63f35d700e8e0d2d1235cc1993f388ad714"
    },
    {
      "title": "What does LLM in AI mean?",
      "block": "geo-basics-2",
      "arm": "placebo",
      "addition": "Keep the explanation of a language model separate from the verification of a company claim. Tokens, context and attention describe how text is represented and generated. They do not by themselves establish that a current product page was consulted or that a statement about an account is supported by the relevant source.\n\nWhen a returned answer provides a link, inspect the dated passage for the particular proposition. Retain the answer's capture date and the reference check date as separate observations. That distinction matters when a historical company allowance differs from what the retained vendor page states. The abbreviation LLM does not resolve the disagreement. The source review concerns the actual claim and the available evidence, while the architectural explanation answers how the model produces text.",
      "addition_words": 125,
      "before_sha256": "3328abbf97f0a5ec354d5edb3df89de5c019ca1861d8459d92cc1ed5afb65ff1",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/what-does-llm-in-ai-mean/",
      "after_text_sha256_lf": "26246d0896d1141a6eb60716e175681b8c503fd0d77574535817e3ba4f933703",
      "addition_sha256_utf8_lf": "84d4c104d28733ebad743007ffaf1fdba9a81b878b78b2d28ba07ff0901ba0fb"
    },
    {
      "title": "What is an AI software shortlist?",
      "block": "geo-basics-3",
      "arm": "treated",
      "addition": "Keep a named option separate from the source attached to it. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) found qualifying own-domain citations in 2,608 of 8,085 named vendor-answer pairs, or 32.3%, in the September 2026 B2B software archive. A pair could be named without a qualifying citation to that vendor's own domain.\n\nThe study states, “A citation describes a source choice.” The percentage is conditional on naming, not a measure of suitability, product quality or purchase. A third-party page may still be cited when the vendor's domain is absent. Read the returned list with those separate observations intact, then inspect the claim and the buyer's requirements. These archived questions do not represent every buying situation or establish that a particular shortlist option deserves selection.",
      "addition_words": 124,
      "before_sha256": "c53fb999b9de07417818e96bedd438b76488e5ee8a21cd1329e3ba586316ebcf",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/what-is-an-ai-software-shortlist/",
      "after_text_sha256_lf": "1902a7aa94b5c8a64bb14accfba7ddac25c789bbe50bd8d604ba1a359ee85766",
      "addition_sha256_utf8_lf": "e90f1949e93f619c1eedb26d4c8259d69973a31a70ff95ff214ca0dee9abc3c1"
    },
    {
      "title": "Does being named by AI prove product quality?",
      "block": "geo-basics-3",
      "arm": "placebo",
      "addition": "Evaluate an appearance and the description attached to it as separate observations. First establish that the product was named; then inspect whether the important claims are supported by dated evidence. A stated feature can justify a narrower description without settling whether the product performs well for the buyer's actual requirements.\n\nKeep both records when a named product is described incorrectly. Correcting the description is a factual task, not a request to certify quality or a conclusion that the product is poor. A comparison should preserve the distinction between stated features, tested functionality and unknown evidence. Use the returned name as a reason to investigate where relevant, then follow the requirement and the available evidence. The frequency of the name alone does not complete that evaluation.",
      "addition_words": 125,
      "before_sha256": "3a90ddd0464008050ffec5a23ed54f6b98ea997b12360331987dc13ac722bb45",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/does-being-named-by-ai-prove-product-quality/",
      "after_text_sha256_lf": "ffe583e00de1335944aa7476414b7b615390f56b416b41ad6cf3bf755911a766",
      "addition_sha256_utf8_lf": "277471f8912f1faf9663f7da6116ab5c533ee9e022608980f2bfa570189f7870"
    },
    {
      "title": "Are failed AI answers the same as a brand absence?",
      "block": "geo-basics-4",
      "arm": "treated",
      "addition": "Preserve the distinction between an excluded attempt and an eligible answer with no citation. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) retained 2,095 eligible target answers in the September 2026 B2B software archive after five target exclusions. Among eligible answers, 39 had empty citation arrays, or 1.9%; the excluded attempts were retained in a separate audit.\n\nThe study states, “A citation describes a source choice.” An empty array records a source outcome for an eligible answer, whereas a failed attempt does not establish a brand absence. Keep both states visible with their own bases. These counts describe this archive and its eligibility rules. They do not make every service failure a valid answer or establish that missing citations and missing brand names are equivalent events.",
      "addition_words": 124,
      "before_sha256": "c0afcfbce3a7cdfc1f52ca34dd9bd60090f48a2d50a6b3f5986b42ee78036972",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/are-failed-ai-answers-the-same-as-a-brand-absence/",
      "after_text_sha256_lf": "91a48c2b3f0a5f39c7eb10ab911ed0bed01fb6229433a24bba5292a74e639ceb",
      "addition_sha256_utf8_lf": "8103d48ee3040c56c2299768aa77bfd65dbaf6ab54b5aec1d5ddc20bba670b30"
    },
    {
      "title": "Does an AI shortlist show what buyers actually purchase?",
      "block": "geo-basics-4",
      "arm": "placebo",
      "addition": "Describe a returned shortlist as a record of the options presented in an answer. Keep later buyer actions in their own records, with the dates and definitions that make those events identifiable. The person may investigate an option, reject the list or make no purchase. The list alone does not distinguish those possibilities.\n\nApply the same boundary when reporting a business outcome. A suitable recommendation still leaves price, requirements, procurement and the person's evaluation unresolved. Preserve visits, self-reports, opportunities and sales as separate events rather than renaming an appearance measure. A later comparison can connect records only through a stated, supported link. If that link is absent, the contribution remains unknown. The answer sample can remain useful while retaining that limit on its interpretation.",
      "addition_words": 124,
      "before_sha256": "42190d7f5932577c5dc377613200e8be4771539c6b3e20aabb3768f8b973e726",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/does-an-ai-shortlist-show-what-buyers-actually-purchase/",
      "after_text_sha256_lf": "4203a671ca67564dfc17ab2cc260edb4ac56e2e9a2ac9d05883199ffa5774142",
      "addition_sha256_utf8_lf": "828d794145bf8642b6a5b09e8eb2651f31777f78c1c1726146f73ee50153e958"
    },
    {
      "title": "Why does the buyer question matter to an AI shortlist?",
      "block": "geo-basics-5",
      "arm": "placebo",
      "addition": "Read the buyer condition before applying a category-wide result to a particular decision. A question about residential builders retains that work type as part of what was measured. Team size, budget and other needs can define a different question, even when both questions sit within the same broad software category.\n\nKeep that wording stable when comparing dates. Replacing the question with an easier or broader one changes the observation being compared. Preserve a missing name as a record of the captured answers under the stated condition, with the uncertainty that accompanies the repeated runs. It does not establish whether the product would suit the particular buyer. The useful interpretation stays attached to the question and the retained observations rather than to the category label alone.",
      "addition_words": 125,
      "before_sha256": "3105b9a375e1404d8655f5b720c77ca794b274732ce5396514ba414ebe0c76ab",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/why-does-the-buyer-question-matter-to-an-ai-shortlist/",
      "after_text_sha256_lf": "7a1aebdd306d9a2c9c1caa49db6aaec3f75af47e2a6ebb0efa2c9624eee0d15d",
      "addition_sha256_utf8_lf": "2e50846fcb0c74b5208ec100f3caf498d651da4a450e060c2038e8747c1942a5"
    },
    {
      "title": "Why can the same AI question produce different names?",
      "block": "geo-basics-5",
      "arm": "treated",
      "addition": "Fix what counts as a name and what counts as a source when comparing repeated answers. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) counted qualifying own-domain citations in 2,608 of 8,085 named vendor-answer pairs, or 32.3%, in the September 2026 B2B software archive. Naming and source selection were separate recorded events.\n\nThe study states, “A citation describes a source choice.” A change in sources therefore need not be a change in the returned vendor set, and neither observation alone identifies why an answer changed. This descriptive analysis does not measure repeat stability or prescribe a sufficient repeat count. Retain the repeated question's wording, dates and separate outcome definitions. The earlier repeat evidence remains the evidence for variation; this later count clarifies an additional boundary in interpreting it.",
      "addition_words": 127,
      "before_sha256": "18591db04b2cb29f5ed78b92810ec28d3df5034c09c10572dd91c85c49700188",
      "url": "https://docs.broadcastwell.com/answers/geo-basics/why-can-the-same-ai-question-produce-different-names/",
      "after_text_sha256_lf": "6d0eeed29d911c743e5f57afca259f7b433127ba9e7863a5a30d6a6327474111",
      "addition_sha256_utf8_lf": "ec4489bf634a94675e5fe0015ce1af8edcfc14f0fc8a48ce0410974e50fa4589"
    },
    {
      "title": "How should a comparison page handle unknown product features?",
      "block": "fixes-1",
      "arm": "placebo",
      "addition": "Keep a comparison cell tied to the precise scope of its supporting page. A statement about a suite, connected tool or different plan may leave the buyer's narrower question unresolved. Review the product, module, region and date before choosing a source-state label, and keep the qualification near the comparison rather than hiding it elsewhere.\n\nUse source silence carefully. If the reviewed page does not state a capability, record that limit instead of treating it as a universal test of the software. An unknown remains a question for further evidence. A later suitable primary source can support an update when the claim and review date change together. Retaining the earlier evidence makes the correction traceable without turning the unresolved cell into an invented feature conclusion.",
      "addition_words": 124,
      "before_sha256": "ad2953b1d9f24a9d72bf27e10f01cecb62d27ffcbf707aae539dd344d2d217ed",
      "url": "https://docs.broadcastwell.com/answers/fixes/how-should-a-comparison-page-handle-unknown-product-features/",
      "after_text_sha256_lf": "9282ac666bbf9267c00685a9ef4fd699ebd377109b61aa0828ad907fb75cab02",
      "addition_sha256_utf8_lf": "d1d9d8cbc68217d02b464db4deacb9a228e14d118c47f1fd0de9bf8dc24858c5"
    },
    {
      "title": "What does publishing llms.txt prove about AI visibility?",
      "block": "fixes-1",
      "arm": "treated",
      "addition": "A published overview file records availability at a particular address, rather than demonstrated use in an answer. Our [registered page study](https://docs.broadcastwell.com/research/studies/optimized-pages-in-ai-answers/) found llms.txt present at 774 of 1,600 target host roots, or 48.4%, during October 2026 retrieval of hosts cited in the September B2B software archive. Another 85 root-file states remained unknown.\n\nThe study states, “Presence does not prove consumption.” This is a descriptive host observation, not a controlled comparison of publishing the file. It cannot establish that a particular service requested the file, used its links or cited a page because of it. Keep the served-file check separate from later request and answer evidence. A successful publication can be verified directly while the file's contribution to a generated answer remains unresolved.",
      "addition_words": 125,
      "before_sha256": "535f72327d0c9d85eb6a282e0f97b38e82726a9dc124df7bc004ddc846b8c26f",
      "url": "https://docs.broadcastwell.com/answers/fixes/what-does-publishing-llms-txt-prove-about-ai-visibility/",
      "after_text_sha256_lf": "81d4d32b138a7a862338a8651377cb391420e23a3a7e11391827960ce15b10f4",
      "addition_sha256_utf8_lf": "30524ef8a7bc72a40446d49d6da5593bf08d95631b33819648119ed457294ed8"
    },
    {
      "title": "What is an llms.txt file?",
      "block": "fixes-2",
      "arm": "placebo",
      "addition": "Give the overview a clear scope before shortening the material it describes. A documentation file should identify the covered part of the site, and a link description should lead to the relevant detailed page. The concise overview and its destinations need to retain the same product boundaries and dates so that shortening does not alter the claim.\n\nKeep the access check separate from that editorial check. An overview does not grant permission to fetch restricted content, and its presence does not establish that a linked page can be retrieved successfully. Follow the destinations and compare the descriptions with what a person reads there. A file can be concise while still making the covered path and the purpose of its selected links clear to the reader.",
      "addition_words": 125,
      "before_sha256": "caed7155b9eaa1760ea3872675a23a4a84d16f6339975ee480aebc4ced3a66ba",
      "url": "https://docs.broadcastwell.com/answers/fixes/what-is-an-llms-txt-file/",
      "after_text_sha256_lf": "99d88013d404eb9886258318d22199f6753aeff87f2a3bbf0af846f080024ec7",
      "addition_sha256_utf8_lf": "d400d5e778f4f3fefc49e616551933d1d130710bc7fe2cb8c37944e7a4d128ee"
    },
    {
      "title": "Is llms.txt actually used?",
      "block": "fixes-2",
      "arm": "treated",
      "addition": "Distinguish finding a file from observing its use. Our [registered page study](https://docs.broadcastwell.com/research/studies/optimized-pages-in-ai-answers/) found llms.txt at 774 of 1,600 target host roots, or 48.4%, during October 2026 retrieval of hosts cited in the September B2B software archive. The root-file states also included 741 absent and 85 unknown cases under the stated response rules.\n\nThe study states, “Presence does not prove consumption.” These data describe file availability on the captured hosts. They do not record whether an answer engine requested an overview or relied on its contents. A request log and an answer observation would address different events and need their own dates and definitions. Preserve the unknown state instead of treating a blocked or ambiguous response as proof that no file exists.",
      "addition_words": 124,
      "before_sha256": "2b7cac22bb8a249f98a1c612243d9caa40967149cc2f8b8285824b1b8dfc35d1",
      "url": "https://docs.broadcastwell.com/answers/fixes/is-llms-txt-actually-used/",
      "after_text_sha256_lf": "0bc3a09ce82ff4dd2add6de787bbecbabde1282c5250a233e52e7e0a25a99a03",
      "addition_sha256_utf8_lf": "40759716f39c97f4f4d1195009b4381c14a8db0b8b60fde4123f5de2b5b841f4"
    },
    {
      "title": "How should I check an llms.txt file?",
      "block": "fixes-3",
      "arm": "placebo",
      "addition": "Record what was checked before calling the file valid. Begin with the public address and its actual response, then inspect the intended text and follow the destinations. A successful retrieval and accepted syntax do not establish that a description matches the current page, the relevant plan or the study the link promises to explain.\n\nKeep those outcomes distinct in the receipt. If a destination is unresolved, leave that state visible rather than hiding it in an overall score. After a deployment, compare the served text with the approved version and revisit changed links. The check then identifies which static facts were confirmed and which remain unresolved. Any later statement about use, citation or recommendation still requires its own observation and stated method.",
      "addition_words": 122,
      "before_sha256": "bc0add69d08349c314e64897059cf43877b61610fd9cde7115921d743fceefa1",
      "url": "https://docs.broadcastwell.com/answers/fixes/how-should-i-check-an-llms-txt-file/",
      "after_text_sha256_lf": "2ff2e51420ea5708518d4a94b611821d870e62d4bd9b88e3de6420115e187d1e",
      "addition_sha256_utf8_lf": "3cd7c07fb377aa053830adad6404b86bf278095b6bceba0ac807cc26c9963122"
    },
    {
      "title": "How should a page distinguish native features from integrations?",
      "block": "fixes-3",
      "arm": "treated",
      "addition": "Check the cited passage for the exact product boundary, even when the answer supplies a link. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) found qualifying own-domain citations in 2,608 of 8,085 named vendor-answer pairs, or 32.3%, in the September 2026 B2B software archive. The matching rule identifies a name and its qualifying domain, not the delivery method of a feature.\n\nThe study states, “A citation describes a source choice.” A source link alone cannot establish whether a capability is native, supplied through an integration or unresolved. Inspect the passage's module, plan and date before using it to fill that comparison. This descriptive count is not a native-feature accuracy test and does not turn an unmatched own-domain citation into proof that a capability is missing.",
      "addition_words": 124,
      "before_sha256": "31362c1f5bc4d6635fecd40dcf62818f90ab4ff4f034b9f66c8a4ea0a26b9f07",
      "url": "https://docs.broadcastwell.com/answers/fixes/how-should-a-page-distinguish-native-features-from-integrations/",
      "after_text_sha256_lf": "5a6073e434859ee37bd4e2a53eb00b03e504de35aa025c4c400ff23e9797d6ad",
      "addition_sha256_utf8_lf": "94a151531a84c56727d99f849c21fc4cd32a745501e74973eb992bd5c90e98c1"
    },
    {
      "title": "How can I generate llms.txt from existing pages?",
      "block": "fixes-4",
      "arm": "treated",
      "addition": "Generate an overview that accurately describes the linked pages, then verify what is publicly served. Our [registered page study](https://docs.broadcastwell.com/research/studies/optimized-pages-in-ai-answers/) found llms.txt at 774 of 1,600 target host roots, or 48.4%, during October 2026 retrieval of hosts cited in the September B2B software archive. Eighty-five root-file states remained unknown under the response rules.\n\nThe study states, “Presence does not prove consumption.” The observation does not compare file generators, validate their summaries or establish a citation benefit from creating a file. Your generation check therefore still needs to compare each description with its destination and retain the page's scope and date. File availability, summary accuracy and later use in an answer remain separate questions. This count supplies context for that boundary, rather than an effectiveness promise.",
      "addition_words": 127,
      "before_sha256": "d130c94ef86ffb2b7eeb5c0bde53cbe6177acb78b67dfad6362752353b3af8ab",
      "url": "https://docs.broadcastwell.com/answers/fixes/how-can-i-generate-llms-txt-from-existing-pages/",
      "after_text_sha256_lf": "72d18299edc452a0a7e52e8b03956a1e5ffeaab7972958dc2b096a15b3c708d5",
      "addition_sha256_utf8_lf": "c291ac7422138bdc1a378d4538f0b7044d13cb9b32206c599839e66344e238b2"
    },
    {
      "title": "How should I choose content for SEO, AEO and GEO?",
      "block": "fixes-4",
      "arm": "placebo",
      "addition": "Let the missing buyer detail define the content task. If the question concerns an accounting integration, preserve the named system, the information that moves and the documented limits as the matters to investigate. A broader article can provide context without resolving that narrower product question. The format should follow the evidence the reader needs.\n\nKeep the underlying question consistent across the SEO, AEO and GEO brief when they concern the same decision. Repeating the same answer under different acronym labels does not supply distinct information. Also preserve the distinction between eligibility and content: a technical access state and an unanswered accounting detail are different problems to establish. The next page change should follow the actual missing fact or technical condition identified in the retained evidence.",
      "addition_words": 125,
      "before_sha256": "815dfc8e6f21bc716b0807b2c8e36cbca584b19733d9dff3f235d6291579b72b",
      "url": "https://docs.broadcastwell.com/answers/fixes/how-should-i-choose-content-for-seo-aeo-and-geo/",
      "after_text_sha256_lf": "d041ebbdcc340a8d13524b1dcfb9a25e2fe119a3ae5c18493ca52dbd0e37395b",
      "addition_sha256_utf8_lf": "37b3d4758da5610406fca11c97c865535d862eb4bd118b693ccebf900f28db3f"
    },
    {
      "title": "How should I request an AI fact correction with a source receipt?",
      "block": "fixes-5",
      "arm": "treated",
      "addition": "Submit evidence for the particular disputed statement, rather than treating a source link as the conclusion. Our [registered citation study](https://docs.broadcastwell.com/research/studies/who-ai-engines-cite/) found qualifying own-domain citations in 2,608 of 8,085 named vendor-answer pairs, or 32.3%, in the September 2026 B2B software archive. This count records source selection under domain-matching rules, not successful factual corrections.\n\nThe study states, “A citation describes a source choice.” A correction receipt must still identify the dated passage, the claim it supports and the remaining uncertainty. These descriptive data do not test whether a correction request will be accepted or change an answer. Keep the original observation and the later response as separate records. A documented request is evidence of the request; any claimed correction needs a subsequent, retained observation.",
      "addition_words": 125,
      "before_sha256": "125c50e46f1ed888e45e64f69d69d10a6118afc6e9ba1ecde2f4baeb209a2a48",
      "url": "https://docs.broadcastwell.com/answers/fixes/how-should-i-request-an-ai-fact-correction-with-a-source-receipt/",
      "after_text_sha256_lf": "4cde1b54647f1375d4ad246dc057bd6a41942b54790695f79bfe5c2899239415",
      "addition_sha256_utf8_lf": "a931c8f1bc9d476b7951a0fc124bfac55dc90e0ed4fe558cbfa53e9cc4105123"
    },
    {
      "title": "What should I include in an llms.txt file?",
      "block": "fixes-5",
      "arm": "placebo",
      "addition": "Choose each link for the task it helps the reader complete. A pricing destination, method page and product page serve different purposes, and the description should make that scope clear. Preserve the relevant date or historical label when shortening an approved description so that the overview does not become more confident than its destination.\n\nKeep the selected set small enough to maintain accurately. Group the essential company, product and research destinations by their purpose rather than reproducing the full navigation. Supporting detail can stay behind the appropriate link when a short heading would conceal its qualifications. Recheck a changed summary against the human-facing page. The file is another publication of the same facts and needs to retain their meaning, scope and permission boundaries.",
      "addition_words": 123,
      "before_sha256": "a690467bf29322226734e042997fe21af2c61f8508a6a37527914d7bf9418d5c",
      "url": "https://docs.broadcastwell.com/answers/fixes/what-should-i-include-in-an-llms-txt-file/",
      "after_text_sha256_lf": "21078f5cd3b8cc5468d8533779bd0025f7b49472f0146c379a51882bdcf20a81",
      "addition_sha256_utf8_lf": "1c2d2a7b16adcd72613d3c7be7b3b402cd62e7a804f52522eef0e1111e8ee7f3"
    }
  ],
  "hash_normalization": "Full source and proposed text: UTF-8 without BOM, LF newlines. The after-text hash excludes the separately prescribed identical release-date metadata update."
}