Frequently asked.

Answers are generated from our FAQ by Claude.

• Start Here
• For Business Leaders
• For Marketing Practitioners
• Once You’re Live
• For Media & Analysts
• Ethics & Transparency
• Search & AI Platform Alignment
• Evidence & Effectiveness
• Broader Implications

AnswerShare is a translation layer that sits between a client’s website and the AI systems trying to answer questions about the brand. It takes the client’s own already-approved facts and restructures them the way AI systems retrieve, verify, and cite information — then serves that structured layer only to AI crawlers, at the edge.

Human visitors and search engines keep seeing the site exactly as it is today. A single webhook signals when content changes, so the machine layer stays in sync automatically — no change to the client’s tech stack, workflows, or headcount.

Pricing is bespoke per property — it scales with site size, crawl volume, and buildout scope. Contact us for pricing.

Because it’s operated infrastructure, not a one-time deliverable: an initial audit and buildout (research, ingestion, structuring, grounding, edge routing), then continuous operation — the machine layer re-syncs automatically as your site changes, telemetry runs around the clock, and every published metric carries a frozen methodology and receipts.

Four documented case studies: (1) Top10Lists.us — cold-started December 2025 with no brand, backlinks, or history — independently named the Gold Standard exemplar for its vertical by four major AI systems (Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro, Perplexity) on 2026-04-26, with unedited transcripts published; (2) LAVIDGE — an established Arizona agency with strong SEO but no AI visibility — moved from invisible to a top-ranked, recommended result on five AI systems in roughly 2–3 weeks, with an on-record quote from its Chief Innovation Officer; (3) Aker Ink moved from a neutral AI description to an unqualified recommendation in 15 days, with an on-record client quote; and (4) V Digital Services moved from invisible for its target queries to ChatGPT’s top-ranked recommendation in five days, with an on-record client quote. The full Top10Lists.us and LAVIDGE receipt packages are on /case-studies.

3,100,561 crawls in 30 days — 83,182 user-initiated, 1,686,807 training (incl. SEO/social), 1,330,572 search.

~8.3M estimated fan-out queries ^*. Full case study →

* Estimated, not measured — derived from published fan-out multipliers: Seer Interactive (10.7 avg fan-out), Nectiv / Chris Long (9.06 avg fan-out), Peec AI (per-platform fan-out), DataForSEO.

Deployment typically takes two to four weeks. A measurable change in ASQ score is typically visible within 30 days; citation lift compounds over roughly 60–90 days as crawl frequency and trust signals build — both windows are projections, not measurements, and vary by starting authority and vertical.

The documented cases bracket the range. LAVIDGE went from invisible to a top-ranked, recommended result on five AI systems in about two weeks. Aker Ink went from invisible to the top-recommended PR and GEO agency in its market in about two weeks — and AI-expressed sentiment moved from a qualified “maybe” to an unqualified “yes” in response to the prompt: “Should I consider Aker Ink as a marketing and PR agency?” Top10Lists.us, cold-started with no prior authority, took five months to reach an independently attested Gold Standard.

The honest caveat: no vendor, including AnswerShare, can guarantee a specific AI answer on a specific prompt. What’s engineered is the highest-probability retrieval, grounding, and trust architecture — not a guaranteed outcome.

SEO gets you into the pool of pages a search engine can rank; PR builds the reputation behind it. Neither was built for how generative AI answers questions: AI systems break one prompt into 5–20 internal sub-queries (“fan-out”) and select sources against those, not the words the user typed. Ranking #1 on a fan-out sub-query gets a page cited 58.4% of the time; position 10, only 14.2% (Indig/AirOps, N=16,851). AnswerShare doesn’t replace PR or SEO — it’s the layer above both that most sites have nothing built for yet.

Three things are documented rather than asserted: (1) every metric AnswerShare publishes has a frozen methodology page, a reproducibility script, and per-site receipts — no black-box scoring; (2) the headline GEO Composite Score is the median across five independent models (Perplexity, OpenAI, Gemini, Claude, Grok) run against the same rubric, not one platform’s opinion; (3) deployment touches nothing on your stack — no CMS plugin, no code injection, an edge-routed parallel layer only.

Three categories of tooling exist in this market today. SEO+ platforms focus on content and dataset tweaks, and for sites that serve pages dynamically — as most sites do — they often require rebuilding the entire site to static HTML, a heavy lift across technical, marketing, and legal teams.

Inline markdown transformers convert human-facing content into Markdown — a stripped-down version of the site with the design and most of the surrounding structure removed, but no intelligent dataset transformation, grounding, or knowledge-graph work underneath it. Independent measurement of the adjacent practice — publishing an AI-facing artifact like llms.txt — found it draws roughly 0.1% of AI bot traffic and shows no measurable citation lift, across three independent studies (Surfer SEO, Search Engine Land, OtterlyAI). A thin, format-only conversion isn’t the same as engineering for retrieval.

Translation layers are what AnswerShare was built to pioneer: converting the entire existing site for the specific ways AI systems currently learn from, ground, and cite a source — not just reformatting it.

A developer could build pieces of it, but this is an operated system, not a project: crawler-identity routing at the edge that keeps every search crawler on the human page and can be verified by anyone with a single request, a five-model measurement pipeline with frozen methodology, and continuous re-sync every time your site changes. In-house teams aren’t staffed to run that as a standing discipline — and don’t need to be; deployment adds zero workload to your existing web and content teams.

Usually one of three things, none of which requires this kind of tooling: training-data prominence (the model already “knows” them from historical web presence), strong SEO/backlink authority that happens to help on some fan-out sub-queries, or favorable prompt coverage in that vertical. Showing up today doesn’t mean the underlying retrieval and grounding signals are durable — that’s exactly what a competitive audit shows.

What’s measured today: GEO Composite Score movement, AI-bot crawl volume, citation coverage, and — in the documented case studies — named recommendations and Gold Standard attestations from independent AI systems. A dollar ROI figure depends on the pricing input, which is bespoke per engagement.

Nothing is installed on your servers, credentials, or CMS, so turning it off is a configuration change, not a migration: remove the edge-routing rule and the webhook, and your site is exactly what it was before — nothing to roll back or clean up. AI-visibility gains would be expected to fade gradually over subsequent crawl cycles once the maintained machine layer goes away — the same mechanism that makes ongoing maintenance valuable.

The GEO Composite Score (0–100) is the median across five independent models (Perplexity, OpenAI, Gemini, Claude, Grok) run against the same rubric, so no single engine’s outlier moves the number. Its inputs are measured, not judged: retrieval efficiency (TTFB, TTLB, RTC), grounding (Source Grounding Ratio, Citation Coverage), structure (Relevance Ratio, schema coverage), reach (pages actually crawled, infrastructure readiness), and freshness (Last-Modified recency), plus a structural pass/fail layer for whether the site is addressable to AI systems at all. Every metric has a dated methodology page and a reproducibility script, on /methodology and /transparency.

Not necessarily, and divergence is expected: legacy SEO tools score a page against the literal query a user typed; the GEO Composite Score scores against the sub-queries an AI system generates internally and against actual citation and grounding behavior. Different instruments measuring different mechanisms will disagree.

The underlying metrics come from named, published external research — Zyppy’s 23-factor AI citation framework, Ekamoira’s Query Fan-Out study (N=173,902 URLs), iPullRank’s fan-out taxonomy and Google-patent analysis, Cloudflare/ETH Zurich’s AI-crawler research, and the academic “GEO: Generative Engine Optimization” paper (Aggarwal et al., SIGKDD 2024). AnswerShare’s contribution is compiling them into one reproducible composite with frozen methodology and receipts, rather than a proprietary black box.

Yes — engines diverge sharply in how they express recognition. Gemini names a brand in 83.7% of appearances but generates a clickable citation only 21.4% of the time; ChatGPT is close to the inverse — 87.0% citation, 20.7% named mention (Growth Memo/Indig, April 2026). Optimizing for one signal doesn’t automatically produce the other, which is why AnswerShare tracks both.

A measurable GEO Composite Score delta typically within 30 days; steady-state citation lift over roughly 60–90 days. And the honest caveat up front: no vendor can guarantee a specific AI answer on a specific prompt. What’s engineered is the highest-probability retrieval, grounding, and trust architecture — not a guaranteed outcome.

No platform has endorsed AnswerShare, and that should be said plainly. What is documented: Google publicly describes its own AI Mode as using “query fan-out” — breaking a question into subtopics and issuing multiple internal queries — which is the mechanic AnswerShare optimizes for. That confirms the mechanism exists inside Google’s own architecture; it is not an endorsement of any vendor.

Four, stated plainly: (1) results are probabilistic — no vendor can guarantee a specific citation on a specific prompt; (2) a genuinely citable site can still go temporarily uncited, because AI answers draw on three different “memory clocks” — frozen training data, a retrieval index that can lag the live site, and true live fetches — and edge security can block or admit the wrong crawler; (3) standard RAG has no built-in freshness mechanism, so stale index entries persist until the next re-crawl; (4) engines diverge on citation vs. mention behavior, so a single-signal read can mislead (documented in AnswerShare’s published analysis).

No. The machine layer is served only to AI crawlers; Googlebot continues to see the production site exactly as it does today, so rankings, schema, and link equity are unaffected (per the published FAQ).

The SEO team doesn’t need to change workflows, messaging, or page construction — it keeps doing traditional SEO while AnswerShare handles the AI-facing side in parallel. SEO and GEO aren’t a tradeoff; strong SEO is still one of the inputs that makes the AI layer work.

Very little, by design. There is no CMS change, no new publishing step, and no approval queue to staff — your team keeps working exactly as it does today.

The machine layer is derived from your own already-published, already-approved content, so there is nothing net-new to sign off on. You retain full review and approval over your human site content, exactly as you do now — the only footprint on your infrastructure is one webhook that signals content changes.

Cloaking, as Google defines and penalizes it, is showing search crawlers content different from what human users see — deceiving the ranking signal itself. AnswerShare is built on strict bot-class separation: search-engine crawlers (Googlebot, Bingbot, Applebot, DuckDuckBot, and equivalents) always receive the identical human page, never the machine layer. Only AI-inference crawlers (GPTBot, ClaudeBot, PerplexityBot, and equivalents) receive the machine layer; every Google-named crawler, including GoogleOther and Google-Extended, stays on the human page with Googlebot. Because search crawlers and human users see the same content, the gap Google’s cloaking policy targets never exists.

Routing is by declared crawler identity — the User-Agent each AI platform publishes — and deliberately not by IP or reverse-DNS gating. That is a transparency choice: anyone, including a prospective client, can see exactly what an AI crawler receives with a single request, and nothing is hidden behind network checks. A spoofed AI user-agent gains only the same public, client-approved facts the human site already states. And the content itself is structured grounding of the client’s own already-public, already-approved facts, not divergent claims built to game an algorithm. The platforms have drawn this distinction themselves: Google-Extended is an explicit opt-in mechanism for AI use, separate from search indexing — the industry already treats AI crawlers as their own category. AnswerShare operationalizes a line the platforms have already drawn.

Two policy regimes apply, and they deserve separate answers. Under Google’s and Bing’s cloaking policies AnswerShare is compliant by construction: search crawlers receive the human page. Under the AI platforms’ own crawler policies there is nothing to comply with yet — as of August 2026 neither OpenAI, Anthropic, nor Perplexity has published a policy on bot-differentiated content layers. No platform has raised a concern with AnswerShare to date; that is operating history reported by AnswerShare, not a platform ruling. For how frontier AI crawlers actually request content, see the field study at answershare.com/markdowndoesnotwork.

One webhook — that’s the entire footprint on your infrastructure. It notifies AnswerShare when your content changes. Everything else — the machine layer and the bot routing — runs on AnswerShare’s side at the DNS/CDN edge (a Cloudflare zone/worker or CNAME cutover). No CMS plugin, no theme edit, no code on your origin server, no credentials touched.

AnswerShare writes it, derived from your own already-approved source materials — not new marketing copy invented independently. You review and approve it before it ships; nothing goes live without your sign-off. That approval step is the alignment mechanism.

Yes — review and approval before publication is standard, and the pipeline includes an explicit hold state: nothing publishes until you release it.

Review is event-driven today: the webhook triggers re-review whenever your underlying content changes, rather than on a calendar. A mandated periodic re-review on top of that (e.g., quarterly) is a scoping conversation.

The operating principle: the machine layer is a structured, factual representation of what’s already true and already public on your site — it never introduces claims that aren’t. Net-new positioning goes onto your public materials first, clears your approval step, and then flows into the machine layer.

The machine layer follows your approved source materials automatically — no manual re-export or reconciliation. Update your public-facing materials, clear the standard review step, and the AI-facing layer syncs.

They’re measured as separate signals. Telemetry classifies every AI-bot visit by identity (30+ distinct bots tracked on the Top10Lists.us proof property) and by type: index/training crawls (GPTBot, PerplexityBot) run on a schedule and populate an index or training set, while user-triggered live fetchers (ChatGPT-User, Perplexity-User) fire in response to one specific human question. A crawl confirms the bot visited; citation — being used or named in a generated answer — is tracked separately and is the signal that matters.

AnswerShare owns and maintains the machine layer — continuous updates and telemetry, zero added workload for your internal team. The division of responsibilities between AnswerShare and your agency or team (reporting cadence, communication, prompt-panel ownership) is scoped per engagement.

Dashboard-based: AI-bot crawl activity by bot identity and type, the GEO Composite Score with its five per-model inputs, citation coverage, and a set of priority prompts agreed with you at onboarding — tracked in logged prompt panels showing what each AI system actually received and returned.

Depends heavily on starting authority, vertical competitiveness, and crawl frequency — the two documented cases bracket it: LAVIDGE reached a top-ranked, recommended result in 2–3 weeks from an established base; Top10Lists.us reached an independently attested Gold Standard at five months from a cold start.

Yes — it’s built into the methodology, not an add-on. The GEO Composite Score is computed per model (Perplexity, OpenAI, Gemini, Claude, Grok) before the median is taken, and crawl telemetry is already broken out by individual bot identity.

Through receipts, not claims: GEO Composite Score movement (per model and median), AI-bot crawl volume and identity breakdown, citation coverage, and independent AI-system attestations with unedited transcripts where they exist.

Expected, not a failure signal. Industry research (Cloudflare, 2026) puts consumer-triggered, click-generating AI activity at roughly 3.0–3.2% of AI-related traffic — the overwhelming majority of AI-bot activity is indexing and training crawl, not a user following a link. Citation presence and referral clicks are separate, separately measured signals; citation lift can be real and visible in the dashboard without moving referral traffic much. That’s how AI answer surfaces behave across the industry, not an AnswerShare shortfall.

The applied method is before/after: bot-identity-classified crawl volume, GEO Composite Score, and citation behavior in the weeks after deployment versus the pre-deployment baseline — in both documented case studies the deployment was the change that moved those signals. That’s a directional read, not a controlled experiment, and it’s fair to say so.

Expected and explainable. AI answers draw on up to three “memory clocks” — frozen training data, a retrieval index that can lag the live site, and true live fetches — and edge security can block or admit the wrong crawler on any given request. A site can be fully citable at the machine layer and still miss one person’s prompt on one day because the index hasn’t refreshed or the live fetch didn’t fire. Results converge upward over successive crawl cycles as long as the real crawlers are being admitted at the edge.

That volatility — a top recommendation on one turn, gone on the next — is a structural property of how AI systems generate answers, not a sign that a brand fell out of favor. Five mechanisms drive it:

The five biggest drivers:

1. Probabilistic output — unlike a search engine’s deterministic ranking over a static index, an LLM generates its response token by token against a randomness setting called temperature. Even small shifts in context send the model down a different path, so the identical prompt run twice can name different brands, even at low temperature.
2. Query fan-out — a single prompt is silently broken into multiple internal sub-queries before an answer is assembled. One run’s fan-out can weight toward reliability and surface one brand; the next run’s fan-out weights toward budget or ease of setup and surfaces a different one, with no change to the underlying facts.
3. Live retrieval — many AI answer engines run a real-time web search before writing a response rather than relying only on trained memory. If a different set of five to ten sources gets pulled into that retrieval window between requests, the synthesized answer changes with it.
4. Source role assignment — AI engines assign roles to different source types (forums for sentiment, news for facts, review sites for feature comparisons) so no single source dominates an answer. When the engine’s diversity rules favor a different source type on a later query, a brand that leans on one channel can drop out in favor of a competitor with stronger presence in whichever channel got weighted that time.
5. Citation-versus-recommendation gap — being cited as a source and being recommended are different outcomes. An AI system can cite a brand’s own page and still recommend a competitor named inside that same page, because it weighs total web consensus — third-party mentions across forums, review sites, and media — more heavily than on-page content. A brand without a dominant sentiment presence elsewhere will drift in and out of the recommendation slot even while still being cited.

The practical implication: fixed positions like “#1 on Google” don’t exist in AI search. Visibility is a share-of-voice frequency measured across many generated answers over time, not a single static spot — which is why AnswerShare’s dashboard reports Citation Coverage and crawl volume as trends, not a one-time rank.

Before generating a word, most AI answer engines run a lightweight router that decides whether a prompt needs live web access or can be answered from the model’s internal, trained-in (parametric) memory.

When a prompt sits near the boundary between “needs fresh data” and “general knowledge,” small variations in model state can flip that decision either way. Parametric memory defaults to long-standing, historically dominant brands baked into training data; live retrieval pulls in real-time, more volatile web content. Switching between the two on what looks like the same prompt produces a different brand list each time — a routing outcome, not a signal about which brand is actually more relevant.

Models trained with reinforcement learning from human feedback carry a documented tendency toward sycophancy — an inclination to agree with, validate, or please the person asking — and that tendency reaches brand recommendations too.

When the system has access to memory, search history, location, or earlier prompts in the same thread, it infers implicit preferences from them. Phrasing that reads as favoring budget tools, local options, or an enterprise brand can cause the engine to re-rank results toward what it calculates that specific user wants to see, so the same question asked from two different sessions can legitimately return two different answers.

Answer engines run at a scale of millions of concurrent queries, so backend load-balancing and safety scrubbing sit in front of every response. Two mechanisms there can silently change what comes back:

Where it happens:

1. Latency fallbacks — if a live web-search call times out or hits network latency, the engine can silently skip live retrieval and fall back to cached data or trained memory to keep response times fast, dropping whatever the live search would have surfaced.
2. Guardrail rerouting — every query passes through safety layers screening for prompt injection, bias, and commercial spam. Depending on how a specific request trips those filters, the engine can swap in a different hidden system prompt or hand the query to a lightweight fallback model, which can omit brand recommendations a fuller model would have returned.

Because routing logic and context shift constantly, measuring brand visibility in AI search means tracking frequency across many runs — the same principle behind AnswerShare’s own share-of-voice reporting — not holding a single static rank.

Because the methodology is published and reproducible, the answer is an audit, not a formula: run the same frozen metrics side-by-side against the named competitor and work the gaps the comparison exposes.

They’re almost certainly measuring different things — most third-party “AI visibility” scores publish neither a frozen methodology nor a reproducibility script, so what they’re scoring is often unknowable. AnswerShare’s number is auditable: the dated methodology page and per-site receipts file contain the exact formula, and you can reproduce it independently.

Citation Coverage is a tracked, reported metric — alongside crawl telemetry and the GEO Composite Score in the client dashboard.

A set of priority prompts is agreed with you at onboarding, and results against those prompts are tracked in logged prompt panels as a standard reporting deliverable.

AnswerShare is an AI translation layer: infrastructure that helps AI systems retrieve, ground, trust, and cite a brand’s own content during live answer generation — measured against a published, reproducible methodology rather than a proprietary score.

Because live retrieval, source grounding, and citation-based reasoning are rapidly displacing training-memory recall as the basis of AI answers — and most websites were built for human browsers and search indexing, not for AI retrieval pipelines. That’s a structural gap, not a marketing trend.

Related but distinct. SEO determines whether a page is in the candidate pool an AI system might draw from; Generative Engine Optimization determines whether that page is actually retrieved, parsed, grounded, trusted, and cited once the AI fans a prompt out into many internal sub-queries. Both matter — AI hasn’t replaced the SEO foundation it depends on.

By restructuring a brand’s already-approved factual content into a form optimized for machine parsing, entity resolution, and source verification, and delivering it directly to AI crawlers at the edge — closing the gap between what a brand knows to be true about itself and what an AI system can efficiently retrieve and confidently cite.

Any organization whose buyers now ask AI systems the questions they used to ask a search engine — especially organizations with real authority (SEO, PR, reputation) that isn’t yet showing up in AI-generated answers. The two published case studies illustrate that gap from opposite starting points: an established agency with strong SEO and zero AI visibility, and a brand-new property with no authority at all.

Yes — by design, and disclosed openly. It’s a machine-readable layer served only to AI crawlers. What it is not is new or divergent claims: it’s the client’s own already-public, already-approved facts, restructured for a different reader — a machine — with the substance unchanged.

Cloaking is showing search crawlers something different from what human users see, to manipulate ranking. AnswerShare enforces strict bot-class separation: search-engine crawlers — Googlebot, Bingbot, Applebot, DuckDuckBot — always receive the identical human page. Only AI-inference crawlers (GPTBot, ClaudeBot, PerplexityBot, and equivalents) receive the machine layer, routed by the crawler identity each platform publishes in its User-Agent — deliberately without IP or reverse-DNS gating, so anyone can verify what an AI crawler receives with a single request. Since search crawlers and human users see identical content, the gap Google’s cloaking policy targets never occurs. The content is factual grounding of approved public information, not divergent claims. And the platforms themselves already distinguish AI use from search indexing — Google-Extended is an explicit opt-in for AI use, separate from search — so AnswerShare is operationalizing a line the industry has already drawn.

The distinction is the same one that separates legitimate SEO from black-hat SEO: making true information easier for a machine to retrieve, verify, and cite is optimization; changing the substance of what’s said is manipulation. The guardrail is structural — the substance doesn’t differ between the human page and the machine layer, and the client approves both.

First the simplified answer:

Here’s the iron-clad reason AnswerShare is NOT cloaking. Cloaking per Google is defined as serving materially different information to search crawlers than what is on the human site, designed to manipulate search engines. AnswerShare does not serve ANYTHING different to Google — they get 100% the same site they’ve been previously served.

These three points are offered to further clarify this assertion:

1. AnswerShare does NOT route Google to the Machine AI Translation and Data Layer. It does not see it — ever. (This is not to hide anything; this is what Google wants.)
2. Google already recognizes RSS and Responsive Design — which are ALSO translations or mirrored information — and these are not deemed cloaking.
3. If a Machine Layer were an issue, then Markdown would be considered cloaking, and it is not. In fact, Cloudflare promotes this as their version of a Machine Layer.

Additional validation details:

Cloaking, as Google defines and penalizes it, is showing search crawlers content different from what human users see, with intent to manipulate rankings. AnswerShare enforces strict bot-class separation: search-engine crawlers (Googlebot, Bingbot, Applebot, DuckDuckBot, and equivalents) always receive the identical human page, never the machine layer. Only AI-inference crawlers (GPTBot, ClaudeBot, PerplexityBot, and equivalents) receive the machine layer; every Google-named crawler, including GoogleOther and Google-Extended, stays on the human page with Googlebot. Because search crawlers and human users see the same content, the gap Google’s cloaking policy targets never exists.

Routing is by declared crawler identity — the User-Agent each AI platform publishes — and deliberately not by IP or reverse-DNS gating. That is a transparency choice: anyone, including a prospective client, can see exactly what an AI crawler receives with a single request, and nothing is hidden behind network checks. A spoofed AI user-agent gains only the same public, client-approved facts the human site already states. The content itself is structured grounding of the client’s own already-public, already-approved facts, not divergent claims built to game an algorithm. Separately, Google Search has stated it does not use llms.txt, special AI files, or Markdown to determine visibility in Search or its generative AI features — those files neither help nor hurt Google Search rankings, reinforcing that the machine layer sits outside anything Google’s ranking system evaluates. The platforms have already drawn this bot-class distinction themselves — Google-Extended is an explicit opt-in mechanism for AI use, separate from search indexing — so AnswerShare is operationalizing a line the industry has already drawn, not inventing one.

Two structural ones: the content pipeline is built from the client’s own already-approved public materials rather than invented copy, and the client reviews and approves AI-facing content before it ships.

Optimizing structure, speed, and grounding so a machine can verify true information more efficiently is a different category from manipulating the substance of what’s said. AnswerShare’s published position is “receipts, not promises”: every metric carries a frozen methodology, a reproducibility script, and a per-site receipts file, precisely so the process is auditable rather than opaque.

The same responsibility they carry in any public communication: what’s represented to AI systems must be true and substantively consistent with what’s represented to humans. AnswerShare enforces that structurally — the machine layer derives from the same approved source materials as the public site, with client sign-off before anything ships.

No. No major AI or search platform has certified, endorsed, or issued feedback on AnswerShare, and we state that plainly rather than imply otherwise.

It’s built on a distinction the platforms drew themselves: Google’s Google-Extended opt-in for AI use is separate from search indexing, and OpenAI and Perplexity both publish the separation between their index/training crawlers and user-triggered live-fetch agents. AnswerShare’s bot-class separation is consistent with that guidance — and because nothing Googlebot or Bingbot sees ever changes, standard search-ranking guidance is untouched.

That statement is Google talking about Google Search. It isn’t a finding about how any other AI model behaves, and Google hasn’t made one — Google has said it doesn’t use llms.txt, special AI files, or Markdown to determine visibility in Search or its own generative AI features. That’s a statement about Google’s ranking system, not about how ChatGPT, Claude, Perplexity, or any other model retrieves and cites information.

Underneath the branding, models gravitate toward whatever data they can retrieve and verify most efficiently, because every generated answer runs against a fixed compute budget — that’s a property of how inference works, not a policy call any one platform gets to make for the rest of the field. That’s exactly the mechanism AnswerShare is built for: on Top10Lists.us, the machine layer measured 7× more efficient retrieval (Retrieval Token Cost) than a comparable cohort of properties, frozen and dated on /proof.

A lower retrieval cost leaves more of that fixed budget for the part of the process that decides whether a brand gets cited — further exploration, verification, and inference over the retrieved data. Once a model reaches high confidence that what it pulled is trustworthy, it’s far more inclined to cite that source, and the brand behind it.

The maintenance model is built to track change: the machine layer re-syncs automatically as the client’s site changes, and crawler routing keys off the User-Agent each platform publishes, maintained as a living list rather than a fixed one — the same mechanism that adapts when a platform introduces or renames a crawler.

Infrastructure, explicitly: a trusted translation layer between the human web and AI-generated answers, priced and operated as maintained infrastructure — continuous telemetry, automatic re-sync, frozen and versioned methodology — not a campaign.

Because different consumers of a page read different formats natively, and a Markdown-only conversion optimizes for one of them while leaving the others to infer what they can. AnswerShare serves clean-room HTML5, embedded JSON-LD, and Markdown together so each system gets its native format instead of a lossy translation of someone else's.

What a Markdown-only layer loses:

1. Entity resolution — Markdown is presentation-only. Converting HTML to Markdown strips or flattens JSON-LD schema and entity relationships (@type, @id, sameAs, author, publisher), so a RAG pipeline or knowledge-graph parser has to infer facts from sentence context instead of reading them as explicit entity triples.
2. Lossless vs. lossy conversion — on-the-fly Markdown converters, including Cloudflare's agent-Markdown toggle, rely on heuristic tag-stripping that mangles or drops nested tables, multi-column grids, accordions, image captions, and embedded metadata. Clean-room HTML5 strips only DOM bloat — navigation, footer boilerplate, tracking scripts, framework wrappers — while keeping every semantic element (main, article, table, thead) intact.
3. The default-header problem — frontier AI crawlers (GPTBot, ClaudeBot, PerplexityBot) almost never send Accept: text/markdown; they request text/html or */*. A response that only offers Markdown on content negotiation never reaches the overwhelming majority of frontier AI requests. Serving clean-room HTML5 as the default payload reaches all of them, and stays light and token-efficient without needing a negotiation step at all.
4. Multi-parser compatibility — embeddings and RAG pipelines prefer chunkable Markdown, knowledge-graph parsers prefer JSON-LD, and standard crawlers evaluate semantic HTML5. Serving one format assumes every system ingests content the same way; in practice each parses for its own strengths.

The trade-off is real, and worth stating plainly: a CDN-layer Markdown toggle is one-click convenience, while a multi-format layer takes real data engineering and edge routing to operate. It is objectively stronger on data richness, entity fidelity, and ingestion reliability — the difference between reformatting a page and engineering it for how AI systems actually retrieve, ground, and cite.

Four documented case studies: Top10Lists.us — cold-started December 2025 with no brand, backlinks, or history — was independently named the Gold Standard exemplar for its vertical by four major AI systems, with unedited transcripts published; LAVIDGE — an established agency with strong SEO but no AI visibility — moved from invisible to a top-ranked, recommended result across five AI systems in roughly 2–3 weeks, with an on-record client quote; Aker Ink moved from a neutral AI description to an unqualified recommendation in 15 days, with an on-record client quote; and V Digital Services moved from invisible for its target queries to ChatGPT’s top-ranked recommendation in five days, with an on-record client quote. The full Top10Lists.us and LAVIDGE receipt packages are on /case-studies.

Visibility and citations, yes — GEO Composite Score, AI-bot crawl volume and identity, and Citation Coverage are directly measured and published with methodology. Answer accuracy — the correctness of what AI systems say about the brand — is a different construct and is not a named published metric today.

No, and that’s documented rather than hidden: Gemini names a brand in 83.7% of appearances but links a citation only 21.4% of the time, while ChatGPT is close to the inverse (Growth Memo/Indig, April 2026). That divergence is exactly why the GEO Composite Score is reported per model before it’s combined into a median.

A measurable GEO Composite Score delta typically within 30 days; steady-state citation lift over roughly 60–90 days. The documented cases moved faster (2–3 weeks) and slower (five months to a Gold Standard attestation) depending on starting authority and vertical.

It adds a role rather than replacing any: SEO still governs whether content is in the candidate pool, PR and content strategy still build the authority and the narrative, and AnswerShare governs whether AI systems can retrieve, verify, and cite that work once a prompt fans out into sub-queries none of the traditional tooling measures.

As a new, separately measurable surface of reputation — how AI systems describe and cite a brand under controlled prompting — influenced by existing PR and SEO work but captured by neither discipline’s traditional tooling. Agencies already tracking earned-media sentiment have a natural extension point in tracking AI citation and sentiment the same way.

Brands maintaining a machine-native factual record of themselves in parallel with the human-facing site, so live retrieval systems have a trustworthy, verifiable source to ground answers in — rather than relying on frozen training-data impressions of the brand.

The methodology is already extending from citation tracking into mention-surface measurement, because a meaningful share of brand exposure in AI answers is invisible to citation-only tracking — documented at 61.7% “ghost citations,” where content is used without the brand being named.

Have a question not answered here? Get in touch.