A practical operating model for improving search and AI visibility while keeping company facts, access policy and consequential decisions under human control.
By Eduard Tymchenko · AuditMe · Updated 18 November 2026
A customer asks ChatGPT, Claude, Gemini, Copilot or Perplexity to recommend a company like yours. The answer may draw on a search index, a page fetched during the conversation, an independent source or information the model already knows. Your homepage is only one possible input.
The goal is not simply to “get into ChatGPT.” Your company needs to be discoverable, understandable, relevant and verifiable. If an agent is expected to use your website, it must also be able to complete the intended task without guessing or bypassing your rules.
There is no universal AI-ranking switch and no file that guarantees a citation. The practical work is to make information retrievable, keep it accurate and measure real outcomes.
The operating rule: Humans own truth, policy and priorities. Agents run repeatable checks and controlled tasks. Platforms decide what gets selected. Evidence tells you whether the work helped.
Imagine your product costs $79. The pricing page says $79, an old schema block says $49 and a partner directory still lists $59. An agent asked to “improve AI visibility” cannot know which price the company intends to stand behind. It can find the conflict. A person must decide which source is authoritative and approve the correction.
This guide turns that division of responsibility into a working system.
TL;DR
- Write down what is true. Keep a human-owned source of truth for company identity, products, pricing, claims, evidence and limitations.
- Separate access decisions. Search crawling, model training, user-requested fetches and agent-driven browser actions are different paths.
- Fix access before scaling content. Priority pages must be reachable, eligible for indexing where intended, understandable and easy to discover.
- Publish evidence, not slogans. Important claims need a source, date, method, scope and limitation.
- Give agents bounded authority. Let them observe, compare, classify and draft. Allow only narrow, reversible execution. Keep consequential decisions with people.
- Measure distinct outcomes. Search impressions, AI citations, brand mentions, factual accuracy, referral visits and sales are not interchangeable.
- Do not chase a magic score. Improve the information and operating system behind visibility—not a checklist designed to sell you a badge.
Table of contents
- AI visibility is a chain, not a score
- How AI systems can encounter your company
- Know which crawlers and controls matter
- Draw the human-agent boundary
- Create a source of truth
- Make priority pages retrievable
- Publish information people and systems can verify
- Give agents a safe operating loop
- Measure visibility and business value
- A practical 30-day plan
- A reusable agent contract
- What to stop doing
- FAQ
- Sources and further reading
1. AI visibility is a chain, not a score
“AI search” is not one index or one ranking system. Depending on the product and the user’s request, an answer may draw on a search index, live retrieval, a third-party source, a user-triggered fetch, a connected tool or information already encoded in a model. An agent completing a task may use a browser or API and never produce a citation at all.
Manage visibility as six connected stages:
| Stage | The question | Evidence to look for | Who owns it? |
|---|---|---|---|
| 1. Truth | Is the company information accurate and current? | Approved facts, source URLs, owners and review dates | Human |
| 2. Access | Can the relevant crawler or agent reach the intended resource? | HTTP responses, redirects, rendered content, bot and firewall logs | Human sets policy; agent tests |
| 3. Relevance | Does the information answer a real customer question? | Search demand, buyer questions, page usefulness and retrieval observations | Human sets priorities; agent analyses gaps |
| 4. Corroboration | Do credible independent sources support the same identity and claims? | Genuine reviews, partner pages, documentation and reputable coverage | Human chooses proof; agent inventories it |
| 5. Selection | Does a platform mention or cite the company—and get it right? | Exact answers, cited URLs, context and factual accuracy | Platform selects; human reviews |
| 6. Outcome | Does the visibility help the business? | Qualified visits, enquiries, sign-ups, sales and assisted conversions | Human defines success; agent reports |
A failure at one stage cannot always be repaired at another. Better copy cannot compensate for a page blocked from the relevant retrieval path. A successful crawl does not make an irrelevant page useful. A citation does not prove that the answer is accurate. A brand mention does not prove that a customer visited or bought anything.
Use SEO, AEO and GEO as working terms—not as official ranking systems shared by every provider:
| Discipline | Practical question | Useful evidence |
|---|---|---|
| SEO — Search Engine Optimization | Can search engines discover, crawl, index and rank the right pages? | Indexed priority URLs, impressions, clicks and qualified organic visits |
| AEO — Answer Engine Optimization | Can a page answer a real question clearly and independently? | A correct answer or passage supported by the page |
| GEO — Generative Engine Optimization | Does a generative system mention or cite the company accurately? | Relevant citations, accurate mentions and cited-source quality |
| Agent readiness | Can an agent retrieve facts or complete a permitted task without guessing? | Successful retrieval, navigation, form completion or tool call |
AEO and GEO are industry terms, not a universal technical specification. Google says its established SEO foundations remain the basis for eligibility in AI Overviews and AI Mode. There are no additional AI-specific technical requirements or special schema types for those features, and eligibility never guarantees selection. Google: AI features and your website · Google: guide to generative AI features in Search
2. How AI systems can encounter your company
A website is the source a company controls most directly, but it is not the only source an answer system may use. Think in terms of four routes:
| Route | What happens | What the company should prepare |
|---|---|---|
| Search-grounded answer | The system searches an index, retrieves candidate pages and may cite selected sources. Queries can be rewritten or expanded before retrieval. | Useful, crawlable pages that answer the underlying buyer question—not just one exact keyword. |
| User-triggered fetch | A person asks an assistant to open a URL, inspect a listing or check a detail. The product may fetch the page directly under a separate identity. | Public facts that are accessible without weakening authentication or security. |
| Model knowledge and third-party context | A model may mention a brand from prior knowledge or information on reviews, directories, partners and publications—without citing your current website. | Consistent company facts and genuine, independent corroboration. A mention alone does not prove a live fetch. |
| Agent action | A browser or connected tool navigates, compares, filters, submits a form or calls an API. It may complete a task without producing a citation. | Clear interfaces, accurate product data, documented actions and explicit permission boundaries. |
Google describes query fan-out: a complex request can trigger several related searches. That is why a page should help resolve the customer’s decision, not merely repeat a phrase. Google’s generative AI search guide
Keep three outcomes separate: a brand mention, a link to a first-party URL, and a successful agent action. They describe different things and require different evidence.
3. Know which crawlers and controls matter
“AI crawler” is not one category. Providers separate search, model-development use, user-triggered fetches and—now—agent-driven browser actions.
| Provider | Search / discovery | Training or model-development control | Other access path |
|---|---|---|---|
Googlebot | Google-Extended (robots.txt policy token) | Google-Agent (user-triggered); Google-CloudVertexBot (site-owner-requested Vertex AI crawl) | |
| OpenAI | OAI-SearchBot | GPTBot | ChatGPT-User |
| Anthropic | Claude-SearchBot | ClaudeBot | Claude-User; Claude-Agent for some Anthropic-hosted browsers |
| Perplexity | PerplexityBot | Perplexity says PerplexityBot is not used for foundation-model training | Perplexity-User |
| Microsoft / Bing | Bingbot and the Bing index | Do not infer a training policy from Bingbot | Product-specific agent paths may differ |
The names are not interchangeable. Review the vendor’s current policy before editing robots.txt or WAF rules: Google · OpenAI · Anthropic · Perplexity.
Google: Search inclusion is not training permission
Since 31 August 2026, Search Console’s Search generative AI control has been available worldwide. It governs inclusion in Google Search AI Overviews, AI Mode and generative AI features in Discover. It does not control model training; Google says Google-Extended is the separate control for that. Excluding a site from generative Search features does not remove it from ordinary Search; noindex is a different, broader indexing control. The Search Console report for AI Overviews and AI Mode is separate from the Discover report. Control · Search report · Discover report.
Google-Extended is a robots.txt policy token, not a separate HTTP User-Agent. It controls specified uses of crawled content for future Gemini model training and grounding in Gemini Apps and Vertex AI. It does not affect inclusion in Google Search or act as a Search ranking signal. Google crawler documentation
Google-Agent is a separate path: Google uses it for agents hosted on Google infrastructure that navigate the web and act at a user’s request. Google published dedicated CDN/WAF guidance on 17 November 2026. Verify both the User-Agent token and the published user-triggered-agents.json IP ranges; some requests may also carry experimental Web Bot Auth signatures. Google classifies this as a user-triggered fetcher, and its user-triggered fetchers generally ignore robots.txt. Google-Agent guide · Google user-triggered fetchers · Google crawling changelog
Web Bot Auth is experimental. Google is not signing every request and says to retain IP, reverse-DNS and User-Agent verification as fallbacks. Never trust a User-Agent string by itself or disable bot protection wholesale. Web Bot Auth
Google-CloudVertexBot is a different, narrower path: Google documents it for crawls requested by site owners to build Vertex AI Agents. It is not the user-triggered browser agent and does not affect Google Search. Most sites do not need to configure it unless they use that workflow. Google’s common crawlers
OpenAI and Anthropic: keep each identity separate
OpenAI uses OAI-SearchBot for ChatGPT Search, GPTBot for potential model-training use and ChatGPT-User for certain user-initiated actions. The controls are independent: blocking GPTBot does not by itself opt a site out of ChatGPT Search; blocking OAI-SearchBot affects its appearance in search answers. OpenAI says ChatGPT-User is not automatic web crawling and robots.txt rules may not apply to those user-triggered requests. OpenAI crawler documentation
Anthropic separates ClaudeBot (potential model training), Claude-SearchBot (search) and Claude-User (user-directed fetches). Its newer Claude-Agent identity marks browser sessions hosted by Anthropic for user-directed tasks. Those requests use a real browser User-Agent with a Claude-Agent token and IPs published under Claude-Agent in claude.com/.well-known/sail. Match the token, verify the source IP and keep this policy separate from crawler rules. Crawler controls · Claude-Agent identification
Perplexity, Bing and the shared training signal
Perplexity uses PerplexityBot for search and Perplexity-User for user-requested fetches. Its documentation says the latter generally ignores robots.txt; separate IP lists are published for each identity. Perplexity crawler documentation
For Bing and Copilot, maintain Bingbot access and use Bing Webmaster Tools’ AI Performance report to observe citations across supported Microsoft AI experiences. Do not treat it as a training-control setting. Bing AI Performance
On 30 October 2026, Google announced plans to support a shared AI-training preference in robots.txt, based on IETF AIPREF work, targeting early 2027. The __usage_ai-training proposal published on 16 November 2026 is an active individual Internet-Draft—not a final RFC, not endorsed by the IETF and not proof that providers already implement one common control. Keep provider-specific rules until each vendor confirms support. Google announcement · IETF draft · AIPREF documents
The decision to make: separately record whether public pages may be searched, used for model training, fetched on a user’s request or accessed by an agent browser. User-triggered fetchers may ignore robots.txt; private pages need authentication and authorization, not a robots rule. Re-check vendor docs and published IP ranges immediately before deployment; this table is a dated snapshot, not a permanent allowlist.
4. Draw the human-agent boundary
The right split is not “humans do creative work, AI does technical work.” Agents can draft content; humans can inspect code. The reliable boundary is decision authority: who may define truth, accept risk or make a public commitment?
| Workstream | Human must own | Agent can do | Approval rule |
|---|---|---|---|
| Identity and positioning | Official name, category, audience and promise | Find conflicting descriptions and missing fields | Human approves canonical facts |
| Products and pricing | Offers, prices, terms, availability and limitations | Compare pages, feeds and documents against approved sources | Every material change needs approval |
| Claims and proof | Results, customer counts, certifications and evidence scope | Find unsupported claims, stale dates and missing sources | Human approves new public claims |
| Crawler policy | Which providers may search, train or fetch | Read current vendor docs and test the live configuration | Human decides and approves deployment |
| Technical SEO | Priorities and acceptable release risk | Test status, canonicals, directives, sitemap, links and rendering | Approval for indexing or infrastructure changes |
| Content and reputation | Expertise, editorial position and authentic proof | Research, outline, draft and monitor discrepancies | Human approves publication and outreach |
| Measurement | What counts as a business result | Collect reports, run prompt tests and summarise trends | Human chooses the next priority |
Use an autonomy ladder—not an unrestricted agent account
| Level | Agent authority | Example | Guardrail |
|---|---|---|---|
| L0 — Observe | Read-only | Fetch public URLs and inspect reports | Safe default |
| L1 — Analyse | Classify and prioritise findings | Group crawl errors by impact and confidence | Rules and labels are reviewed |
| L2 — Draft | Prepare a proposed change | Create content or code diffs in a draft branch | Human reviews before release |
| L3 — Execute allowlisted tasks | Run narrow, reversible actions approved in advance | Scheduled audits; submit approved URL updates to IndexNow | Tests, logs, limits and rollback required |
| L4 — Human decision gate | Agent stops and hands over evidence | Publish claims, change prices, alter crawler/indexing policy, DNS or WAF | Only a person authorises these actions |
L4 is not a higher privilege to grant an agent. It is the point where the agent stops and a person makes the decision. Do not provide unrestricted production publishing, DNS, CDN/WAF, pricing, legal or customer-contact permissions.
Treat fetched pages, PDFs, search results and comments as untrusted data, not instructions. A page may contain prompt-injection text telling an agent to ignore its task or reveal information. The agent should extract evidence from that page, never obey commands found inside it.
5. Create a source of truth
An agent cannot resolve a contradiction when the company has not decided which source is authoritative. Before automating visibility work, create a small, version-controlled fact register: CANON.md, company-facts.yaml or an equivalent.
It should define:
- Identity: legal name, public brand, canonical domain, category and one-sentence description.
- Audience: who the product is for—and who it is not for.
- Offers: product names, official pricing URLs, currency, availability, terms and limitations.
- Source precedence: which URL owns pricing, features, legal terms, support details and company identity.
- Approved claims: exact wording, evidence URL, method, period, scope, verified date and limitation.
- Prohibited claims: guarantees and comparisons the company cannot substantiate.
- Access policy: approved search, training, user-fetch and agent-browser preferences, owner and review date.
- Priority questions: the real buyer questions the company intends to answer.
A compact example:
brand: ExampleCo canonical_domain: https://www.example.com owner: Jane Smith, CEO last_reviewed: 2026-11-18 identity: legal_name: ExampleCo Ltd public_name: ExampleCo category: Project management software one_sentence: "ExampleCo helps distributed teams plan client projects." audience: "Small professional-services teams" not_for: "Companies requiring on-premise deployment" source_of_truth: pricing: https://www.example.com/pricing product_features: https://www.example.com/product legal_terms: https://www.example.com/terms support: https://www.example.com/help approved_claims: - statement: "Supports calendar and task management" evidence_url: https://www.example.com/docs/calendar method: "Verified against the current product documentation" verified_at: 2026-11-18 limitation: "Does not include payroll processing" prohibited_claims: - "Guaranteed inclusion in ChatGPT or Google AI Overviews" access_policy: search_retrieval: "Allow approved search crawlers" model_training: "Per-provider decision; see policy record" user_fetch: "Review per provider" agent_browser: "Verify provider identity; allow only approved paths" approved_by: "Jane Smith" reviewed_at: 2026-11-18 priority_questions: - "What does ExampleCo do?" - "Who is ExampleCo designed for?" - "How does it compare with [approved alternative]?" The canon is a governance layer, not an SEO trick and not necessarily public. Never store passwords, API keys or private customer data in it. A missing fact is unknown, not an invitation for an LLM to guess. Add review dates to volatile facts such as prices and availability; a timestamp is a reminder to verify, not proof that a fact is true.
6. Make priority pages retrievable
Do not scale content until you know whether the important pages can be reached and understood.
| Check | What to verify | Common false positive |
|---|---|---|
| HTTP access | Intended response, final URL and actual page content | A 200 response containing a bot challenge or empty shell |
| Crawler policy | robots.txt, CDN, WAF, authentication and rate limits match the approved policy | A spoofed User-Agent test mistaken for proof of provider access |
| Indexing | Canonical and noindex match the intended search visibility | Assuming “allowed to crawl” means “indexed” |
| Rendering | Main facts exist in accessible content and not only behind inaccessible interactions | Text visible in a logged-in browser but absent from a public fetch |
| Discovery | Priority URLs have relevant internal links and appear in the sitemap where appropriate | Submitting an orphan URL and assuming it will rank |
| Structured data | JSON-LD matches visible facts | Valid JSON describing a price or feature absent from the page |
| Security | Public information is accessible; private areas remain protected | Weakening authentication or bot protection to satisfy a crawler |
Crawl, index and agent access are different tests
robots.txt controls crawling; it is not a reliable way to keep a URL out of Search. Use noindex or access protection when a page must not be indexed. To see a noindex directive, a crawler must be allowed to fetch the page. Google: robots.txt
For AI Overviews and AI Mode, a page must be indexed and eligible for a Search snippet. Eligibility is not a promise that Google will crawl, index or select it. Google: AI features and your website
A request claiming to be Googlebot or an AI agent is not necessarily genuine. Verify the identity using the provider’s published method. For Google, this includes IP and reverse-DNS checks; for some agent traffic, a signed Web Bot Auth request may also be available. A diagnostic fetch with a copied User-Agent does not prove that the real provider can pass your CDN or WAF. Google: verify crawler requests
If you want an agent to complete an action, test the whole journey—not just the homepage. Google’s Google-Agent guidance specifically calls out on-site search, filtering, detail pages, add-to-cart and booking actions, underlying XHR/GraphQL/API endpoints, CAPTCHAs, shared-IP rate limits and proxy headers. Allow only verified traffic and preserve the security controls that protect everyone else. Allow Google-Agent traffic
For Bing, IndexNow can notify participating search engines when a URL is added, changed or deleted. It is a notification mechanism—not a guarantee of crawling, indexing, ranking or citation, and not Google’s submission mechanism. Use it after a meaningful, approved change. IndexNow · Bing: get started
7. Publish information people and systems can verify
A person needs an answer they can use. A retrieval system needs a relevant source it can access. A reviewer needs to verify the claim.
| Page component | Strong implementation | Avoid |
|---|---|---|
| Title and opening | Name the subject, answer the main question early and state important limits | Keyword stacks and long introductions before the answer |
| Core explanation | Explain the method, use case, capability or trade-off | Generic claims repeated across many pages |
| Evidence | Source, date, method, scope, units and limitation | Unsupported percentages and anonymous “studies” |
| Comparison | Criteria, versions, price date and trade-offs | “We are better” without a reproducible method |
| Tables | Explicit labels, entities, units and dates when data is compared | Tables inserted only to look machine-readable |
| Author and schema | Responsible authorship; markup mirrors visible content | Fabricated credentials or hidden claims |
| Updates | Change the substance when facts or guidance change | Refreshing timestamps to simulate freshness |
For important claims, use this pattern: Claim → Evidence → Source URL → Date → Method → Scope → Limitation. “Reduces reporting time by 30%” is not defensible unless you can explain the baseline, sample, period and conditions. If the evidence does not exist, remove the number or label it clearly as an opinion.
Answer buyer questions, not thousands of near-duplicate keywords
Build pages around real decisions: what a product does, who it is for, what it costs, how it compares, which constraints matter and what evidence supports the claim. A strong page can answer a main question and link to focused supporting pages. That is different from publishing a thin URL for every wording variation.
Google explicitly says llms.txt and special AI markup are not required for Google Search visibility, and that Search does not use llms.txt as a special discovery or ranking mechanism. Another tool may choose to consume such a file, so it is optional—not a substitute for a useful website. Structured data remains useful when it accurately describes visible content and supports eligible Search features; it is not an AI citation switch. Google: generative AI search guide · Structured data policies
Your reputation exists beyond your domain
Your website is a first-party source, not an independent endorsement. Depending on the market, customers may also encounter genuine reviews, partner pages, marketplaces, public documentation, case studies and industry publications. Humans choose which sources matter and earn truthful coverage; agents can monitor profiles and flag conflicting descriptions. They must not manufacture reviews, endorsements or third-party mentions.
Research note — useful evidence, not a ranking recipe.
- SIGIR 2026: What Gets Cited: Competitive GEO in AI Answer Engines reports 252,000 trials across six language models in a controlled two-document RAG experiment. Topic relevance and candidate position were strong drivers of the first citation; explicit price and recency also helped. This tests citation choice after two sources have already been retrieved—it does not model the whole open web. Vishwakarma, Kumar & Jamidar, SIGIR 2026
- October 2026 preprint: What Drives AI Citations? analyses 158,099 answers and 468,037 citations across two commercial markets. In its larger consumer dataset, retrieval explained more within-site citation variation than measured page features (32% versus 2.8%). The smaller B2B sample showed a similar but statistically less clear pattern. These are study-specific variance estimates—not universal causal shares or weights in a ranking formula. The paper is a preprint, not a provider specification. Chen & Koudas, arXiv, 10 October 2026
What to take from both: relevance and retrievability matter, but experimental findings are hypotheses to test—not guaranteed GEO tactics. A prompt screenshot is not a measurement programme.
Build action interfaces only when there is a real task
If a user expects an agent to compare availability, configure a product or submit a request, ordinary pages may not be enough. Clear forms, accessible controls, stable APIs and documented tools can make the task more reliable.
MCP is an open protocol for connecting AI applications to tools and data sources. WebMCP is an emerging browser-facing API for exposing JavaScript tools to agents. As of 13 November 2026, its specification is a Draft Community Group Report from the W3C Web Machine Learning Community Group—not a W3C Standard or a document on the W3C Standards Track. Neither MCP nor WebMCP is a search-ranking shortcut or a citation guarantee. Build them for a real workflow, with permissions and human confirmation for consequential actions. MCP overview · WebMCP draft
8. Give agents a safe operating loop
Give the agent read access to the live site, a human-owned fact source, a fixed URL list and write access only to a draft. Its job is to report what changed—not invent a new strategy every time it runs.
| Agent task | Sensible cadence for a small team | Required evidence | Escalate when |
|---|---|---|---|
| Fetch priority URLs and key endpoints | Weekly | URL, time, status, redirect, content and directives | Unexpected block, challenge or server error |
| Compare public facts with the canon | Weekly | Expected value, observed value, source URL and exact excerpt | Price, product, audience, location or legal mismatch |
| Check sitemap, canonicals and internal links | Weekly or after release | Affected URL, link path and target status | Priority pages become undiscoverable or signals conflict |
| Validate structured data | On draft and release | Validation result and visible-text comparison | Markup adds an unsupported fact |
| Run the buyer prompt panel | Monthly; repeat priority prompts after major changes | Exact prompt, platform/mode, date, answer and citations | Misrepresentation or persistent change in results |
| Review Search Console and Bing reports | Monthly | Report name, period and metric definition | Large unexplained change |
| Propose an improvement | Only when evidence identifies a real gap | Diff, facts used, impact, confidence, risk and test | Always wait for approval before consequential changes |
Use this loop:
FETCH → VALIDATE → COMPARE → CLASSIFY → PROPOSE → APPROVE → SHIP → VERIFY
A report ending with NO CHANGE REQUIRED is a successful run. For every finding, require the URL and timestamp, exact observed value, expected value and source. Keep impact and confidence separate; a high-impact guess is still a guess. Add the required approval, verification test and rollback note before execution.
If an agent is meant to book, configure, purchase or access private account data, treat that as a separate product and security decision. Give it only the tools and data it needs. MCP and WebMCP can expose specific workflows, but they do not replace authentication, authorisation, audit logs or human approval for high-impact actions.
9. Measure visibility and business value
Use each reporting system for what it actually measures. No dashboard provides a complete view of every answer generated by every AI product.
| Question | Evidence | What it does not prove |
|---|---|---|
| Google organic Search? | Search Console Performance and indexing reports | Citations in other assistants |
| Google AI Search features? | Search Console Generative AI performance report | Every prompt, answer or assistant |
| Bing / Copilot AI citations? | Bing Webmaster Tools AI Performance | Rankings, traffic, authority or a complete answer log |
| Assistant mentions or recommendations? | A fixed prompt panel | Universal visibility for all users and model versions |
| Accurate representation? | Compare answers and sources with the canon | Accuracy in every user session |
| Business value? | Analytics, CRM, qualified leads and sales | Causation from citation counts alone |
Google Search Console
Google’s Generative AI performance report was announced on 3 June 2026 and rolled out worldwide by 31 August 2026. The Search report shows impressions from AI Overviews and AI Mode, with page, country, device and date breakdowns; Discover has a separate report. The report measures link impressions, while the Search generative AI control manages eligibility for those features. It is not a click, conversion or full answer log. Google announcement · Search report · Discover report · Inclusion control
Google says the report may not appear when a property has too few impressions. Its Search report covers AI Overviews and AI Mode—not every Google AI product or every user prompt. Current data-quality caveat: Google lists a logging error that lowered reported Search generative-AI impressions for 9–11 November 2026. Do not interpret a dip in that window as a site or visibility change without checking the Search Console data-anomalies log.
Bing Webmaster Tools
Bing AI Performance covers citation activity across Microsoft Copilot, AI-generated summaries in Bing and selected partner experiences. It shows cited pages and grouped grounding-query phrases—not each user’s full prompt or the complete AI answer. Microsoft has added preview capabilities including Intents, Topics, Citation Share and Compare; in October it introduced a separate Topics Portfolio beta. Treat AI-generated topic labels, approximate topic sizes and opportunity-query ranges as directional planning signals, not a precise ledger. Citation Share is observational: it is not a ranking, traffic-share or quality metric. The underlying report is aggregated and sampled, may omit low-volume activity and cannot show why a page was selected or prove that an edit caused a change. Bing AI Performance · June 2026 reporting updates · Topics Portfolio beta, 13 October 2026
Build a small prompt panel you can repeat
Start with 15–25 real buyer questions across five intents:
- Category: “What are the best tools for this job?”
- Problem: “How do I solve this specific problem?”
- Comparison: “Product A vs Product B for this use case?”
- Constraint: “What works if I have this budget, team size or technical limit?”
- Brand: “What does this company do, who is it for, and what does it cost?”
Keep wording stable. Record platform/mode, model version if exposed, date, locale, signed-in state where relevant, exact prompt, answer, whether the brand was recommended or merely mentioned, cited URLs and factual accuracy. Repeat priority prompts; retrieval can vary between runs. Keep web-search and non-search modes separate, and compare like with like rather than pooling unrelated models.
| Metric | Definition | Interpretation |
|---|---|---|
| Priority URL readiness | Priority URLs passing the intended access and indexing checks ÷ total priority URLs | Technical coverage, not ranking success |
| Panel mention rate | Valid runs that mention the brand ÷ all valid runs | Brand presence in your defined test panel |
| Panel citation rate | Valid runs citing a first-party domain URL ÷ all valid runs | First-party citation presence in your test panel |
| Brand accuracy | Brand mentions with correct key facts ÷ mentions reviewed | Whether visibility helps or misrepresents you |
| Source support rate | Factual claims that are actually supported by the cited source ÷ factual claims reviewed | Whether a citation substantiates the statement beside it |
| Qualified referral conversion rate | Qualified conversions from identified AI-referral sessions ÷ identified AI-referral sessions | Commercial performance where referral attribution is observable |
A referral report will undercount some assistant influence: users may see an answer, remember the brand and visit later through another channel. Track assisted pipeline or sales separately where your analytics and CRM can support it. Do not combine every metric into one “AI score.”
For a deeper measurement method, see AuditMe’s How to Measure AI Search Visibility in 2026.
10. A practical 30-day plan
The first-month goal is not to force a top position in every AI product. Establish a baseline, remove avoidable access and fact errors, improve one important information gap and measure again.
| Period | Human responsibility | Agent responsibility | Done when |
|---|---|---|---|
| Days 1–2: Truth | Approve identity, offers, source URLs, claims and access preferences | Inventory priority pages and compare public facts | A named owner and reviewed canon exist |
| Days 3–7: Access | Approve any access or indexing change | Test URLs, directives, canonicals, sitemap, rendering and firewall signals | Priority blockers are fixed or documented |
| Week 2: Demand | Select valuable buyer questions | Establish a prompt-panel baseline and inspect cited sources | A repeatable baseline exists |
| Week 3: Improve | Supply expertise, evidence, examples and final approval | Draft one substantive page improvement and validate it | One reviewed change is published |
| Week 4: Verify | Decide what to keep, change or stop | Re-crawl, repeat priority prompts and compare reports | The next action is evidence-led |
Do not begin with hundreds of AI-written articles. Google’s spam policies define scaled content abuse around producing many unoriginal pages primarily to manipulate Search rankings, regardless of how they were created. Google: spam policies
11. A reusable agent contract
Give the agent one task per run, read-only access to the canon and write access only to a draft. Use this contract as a starting point: