AI search & automationYour questions, answered.

No. Access can support discovery or retrieval, but systems decide what to select for each question. Relevance, indexing and response context still matter. Treat crawler access as technical eligibility where applicable, not a promise of recommendations, traffic or inclusion.

Next step: Check access and track actual referrals separately from sampled answer checks.

OpenAI describes OAI-SearchBot for search discovery and GPTBot for potential model training. These are different purposes with separate controls. An allow or block decision for one bot should not be assumed to apply to every provider or every form of retrieval.

Next step: Review search and training policies separately using current OpenAI documentation.

Check that relevant public pages are accessible and not inadvertently blocked for supported search crawling. Robots rules, firewalls and bot protections can all affect access. Accessibility does not guarantee selection, and a successful ordinary browser request is not proof a crawler is admitted.

Next step: Ask your web team to inspect bot access and live responses.

This is a content-policy decision distinct from search discoverability. OpenAI provides controls for sites wishing to exclude pages from potential training. Review the stated purpose and your organisation's preferences; a single rule does not control all providers or data sources.

Next step: Document provider-specific search and training access choices.

Robots rules control compliant crawling, not secure access. A blocked URL can still be discovered from links, and blocking access can prevent crawlers reading noindex directives. Private data needs authentication or other proper access controls, not just a robots entry.

Next step: Protect private material and use suitable indexing directives for public exclusions.

Fetching a page, storing information and choosing it for a result are separate stages. A fetched page may remain unindexed or unselected. Neither technical compliance nor repeated requests guarantee serving for a particular query, so diagnose the actual stage involved.

Next step: Investigate crawl and indexing evidence before drawing ranking conclusions.

Google says existing Search best practices remain relevant and there are no additional technical requirements for its AI features. Special files or markup should not be sold as guaranteed access. Clear, useful and accessible content remains important without assuring any placement.

Next step: Prioritise ordinary Search eligibility and customer usefulness.

No. Markup can clarify meaning and support eligibility for documented features, but systems retain selection control. It must match visible content accurately. Do not infer higher rankings or generative citations simply because a validation tool reports no errors.

Next step: Validate visible-content alignment and measure actual observed results.

FAQ markup is not a reliable promise of citations or recommendations. Structured data has documented uses and limits; relevance and evidence matter beyond its presence. Publishing many generic questions solely for markup adds little value and can create quality problems.

Next step: Use suitable markup for genuine content without making visibility guarantees.

Observed referral visits and relevant tracking parameters can identify some clicked traffic. They cannot measure every mention or exposure, and later direct visits may lack attribution. Keep self-reported enquiry sources separate from tracked referrals to avoid overstating what the evidence shows.

Next step: Report known referrals and optional discovery-source answers as separate measures.

Search-enabled products can retrieve current sources, while model-only responses may rely on learned information. Product settings and context affect behaviour. A model API response is not automatically evidence of what public assistant users see, and citations still need verification.

Next step: Label the product, method and date for every answer check.

A test records one response in a particular context, not a stable universal ranking. Products, retrieval, accounts and time can change results. Repeatable monitoring can reveal patterns, but should not be presented as guaranteed recommendations for all users.

Next step: Store prompts, dates, products and actual citations as sampled observations.

No. Providers publish their own user agents and policies, and rules have specific scope. Robots signals are not universal privacy or training controls. Check each provider's official guidance instead of copying a list and assuming it manages every possible access path.

Next step: Maintain a crawler inventory with stated purposes and intended permissions.

Retrieval finds information for a response; training changes or informs a model's learned patterns. A page being accessed for one purpose does not prove use for the other. Provider definitions and controls vary, so avoid treating all bot traffic as equivalent.

Next step: Separate access policies and reporting by the documented crawler purpose.

Volume alone is not useful. Google warns against scaled content that adds little value, regardless of production method. A focused library can help when questions are distinct, answers accurate and visitors gain practical guidance. Review overlap and evidence rather than filling a numerical quota.

Next step: Publish only distinct, useful answers and maintain an editorial review process.

Generated content can invent facts, omit limits or misread sources. Review the visible answer, metadata and any structured claims against evidence. An AI draft is not independent expert approval; important technical or legal statements may require a qualified reviewer.

Next step: Check substantive claims and dates before publication.

No. It is an optional convention, not a universal control recognised identically by every assistant. It can provide an organised guide where supported, but cannot replace crawl access, accurate content or relevant evidence. Do not present it as a ranking mechanism.

Next step: Maintain useful links without claiming that the file ensures citations.

A citation helps readers inspect a source but may not support every associated sentence. Search results can be incomplete or outdated. Check material claims directly, especially numbers, dates and consequential advice, instead of treating a link as automatic verification.

Next step: Open cited sources and compare the exact claim with the evidence.

Start with bounded tasks, reliable inputs, permission controls and visible failure handling. Drafting or organising can be automated more safely than high-impact publication or customer decisions without review. Define human approval and recovery before expanding scope.

Next step: Pilot one task with a named reviewer and a stop condition.

Separate observed visits, self-reported enquiries, sampled citations and inferred exposure. State the date range, product and method. Do not merge model checks into claims about public assistant rankings; the evidence types answer different questions.

Next step: Use distinct report labels for traffic, enquiries and sampled answers.

Test relevant questions and record the product, model, date and whether search was enabled. Answers vary and a model API response is not identical to the public Claude experience. Combine tests with genuine referral and enquiry evidence rather than claiming a fixed ranking.

Next step: Record a repeatable set of customer questions and the testing conditions.

Use representative questions and distinguish Gemini app answers from API model checks and Google Search features. Their sources and behaviour can differ. A single mention is evidence of that test, not proof that every customer receives the same recommendation.

Next step: Record the exact experience, question and observed source links.

AI Overviews appear within eligible Google search experiences, while chatbots may use different retrieval and generation systems. Availability varies by query and context. Neither should be treated as a uniform result that every user sees for the same wording.

Next step: Assess actual relevant search experiences separately from model prompt tests.

AI Mode is a Google search experience with conversational capabilities and generated responses where available. Its behaviour and availability evolve. Apply useful public content and search fundamentals rather than assuming a separate tag guarantees inclusion.

Next step: Review Google's current AI search guidance and your public page accessibility.

Review relevant questions, cited pages and the product context, recording dates and conditions. Sources and recommendations can change between runs. A citation may reference a factual resource without recommending the business commercially.

Next step: Separate citation evidence from direct business recommendation evidence.

Generative engine optimisation is an industry term for improving discoverability and usefulness in generated answers. It is not a single official platform programme with guaranteed rules. Useful content, accurate facts and accessible public pages remain practical foundations.

Next step: Identify specific discoverability problems rather than buying a guarantee of AI rankings.

The labels overlap: both can involve clear answers, accessible content and credible evidence. Generated answer systems add retrieval and selection behaviour that varies by provider. A separate label does not prove a special technique is required or effective.

Next step: Focus on a documented customer question and the quality of its public answer.

Search Console reports Google search performance according to its current definitions, including treatment of eligible AI features. Available breakdowns may not isolate every generated experience. Check current documentation before labelling a general search total as AI-only traffic.

Next step: Document which AI-related information the current report actually separates.

Record the observed source or the customer's optional self-reported answer and keep these distinct. Someone may discover a business in an assistant and later visit directly. Do not overwrite their explanation with an inferred source solely because analytics cannot confirm it.

Next step: Add an optional discovery-source question and preserve the answer clearly.

Show confirmed enquiries with known AI referral or self-reported discovery, defining the method used. Attribution can be incomplete and does not establish causation for every sale. Avoid counting bot visits or prompt-test mentions as customer leads.

Next step: Report measured referrals and self-reported enquiries as separate evidence categories.

It shows a recorded request attributed to a crawler under the identification method used. It does not prove indexing, training use, citation or recommendation. User-agent claims alone may be unreliable and should be validated where possible.

Next step: Label crawler observations separately from mentions and customer visits.

Use the provider's documented methods where available, rather than trusting a user-agent string alone. Verification methods can change and may involve published network information. Record uncertainty when requests cannot be reliably attributed.

Next step: Compare the identification approach with current official provider guidance.

Review each provider's documented crawler purposes and apply deliberate choices for public retrieval and training where supported. Controls vary and do not cover all systems. Robots preferences are not privacy protections for information that must remain confidential.

Next step: Create a provider-by-provider policy based on current official documentation.

State accurate services, identity, contact routes and genuine locations clearly in visible public content. Keep supporting pages consistent and distinguish coverage from offices. Clarity helps readers and retrieval systems understand facts but does not guarantee selection.

Next step: Audit your published business facts against current operations.

Retrieval systems may encounter several sources describing a business. Accurate listings help reduce conflicting information and give people verification routes. They are not guaranteed inputs for any particular assistant and cannot replace useful information on the business's own website.

Next step: Correct important legitimate profiles and document their current descriptions.

Genuine reviews provide third-party customer evidence that some systems or readers may encounter. Their influence varies and cannot be guaranteed. Follow review platform rules and avoid fake feedback, incentives or claims that a review campaign guarantees recommendations.

Next step: Request honest feedback through appropriate channels and maintain accurate profiles.

Describe the real problem, contribution and verified outcome with scope and permission. Specific evidence is more useful than generic superiority claims. Do not invent figures or assume an assistant will cite a case study because it exists.

Next step: Publish one approved example with verifiable context and limitations.

Use sources that directly support factual claims and give readers a clear route to verify them. Prefer current primary guidance where suitable. Merely attaching many references does not make inaccurate reasoning trustworthy or secure an AI citation.

Next step: Check each important factual statement against its linked evidence.

First-hand expertise draws on actual experience, observed work or qualified knowledge relevant to the subject. Explain its scope and supporting evidence. A generic claim of expertise or an invented personal story does not provide the same value.

Next step: Add a genuine example and clarify what you directly observed.

Accurate transcripts make spoken information available in text and can improve accessibility. Their use by retrieval systems varies. They should preserve meaning, identify relevant context and avoid publishing confidential discussion or unsupported claims.

Next step: Provide a reviewed public transcript for an informative business video.

Relevant images and accurate surrounding descriptions can help communicate products, work or places. Capabilities differ across systems, and image presence does not guarantee understanding or citation. Do not rely on decorative visuals to convey essential facts unavailable in text.

Next step: Add clear visible explanations for important visual evidence.

Choose a structure that helps readers navigate distinct needs without repeating thin answers. A searchable bank with substantive topic views can preserve context and stable links. Hundreds of near-identical pages are not a shortcut to citations or rankings.

Next step: Group related questions into maintained, useful topics.

Answer genuine questions directly, explain relevant limits and provide a practical next step. Avoid padding, unsupported certainty and repeated answers with changed wording. Value depends on the usefulness and accuracy of each entry, not the size of the library.

Next step: Review each question for a distinct customer need and a complete answer.

Assign content owners, review changing guidance and correct dated claims. Preserve stable links where practical so shared references continue working. A published review date should reflect meaningful checking rather than an automatic timestamp.

Next step: Schedule reviews for time-sensitive answers and track substantive corrections.

Record the inaccurate answer and its sources where shown, then correct inaccurate owned content and legitimate listings. Use available provider feedback routes. A correction cannot guarantee immediate change across models, cached content or every future response.

Next step: Identify the mistaken fact and update its authoritative public source.

Determine the product context, date and whether the answer retrieved current sources. A model may rely on older information or repeat an unreliable source. Improve the public evidence and record changes without promising a particular refresh schedule.

Next step: Compare the answer with current primary business facts and cited pages.

AI share of voice is a defined comparison of mentions or recommendations within a chosen question set and testing method. It is not a universal market statistic. State the sample, providers and variability so percentages do not imply comprehensive measurement.

Next step: Define the question set and counting rules before comparing competitors.

Use genuine customer alternatives and relevant question context, then review suggested domains critically. Generated answers can include publishers or unrelated businesses. A useful comparison should assess fit and evidence rather than automatically importing every mentioned name.

Next step: Validate the competitor list with real service and audience overlap.

Check whether the cited page supports the answer and whether the reference is current and authoritative for that claim. A citation can be unrelated, incomplete or wrong. Mention volume should not obscure misleading descriptions or weak evidence.

Next step: Open and verify citations from a sample of important answers.

No legitimate supplier can guarantee how independent systems choose organic recommendations. Paid placements or partnerships, where available, are different from organic selection and should be described accurately. Be cautious of promises lacking a transparent measurement method.

Next step: Ask suppliers to distinguish paid exposure from organic mentions and guarantees.

Highlight verified progress and useful actions without concealing losses, missing data or uncertainty. Separate citations, recommendations and customer outcomes. Positive framing should explain practical improvements rather than invent success when checks do not show it.

Next step: Include evidence, limitations and next actions for each reported result.

Avoid submitting secrets, unnecessary personal data or confidential material without an approved purpose and appropriate safeguards. Provider terms and settings differ. A public chatbot interface should not be assumed equivalent to an approved controlled business environment.

Next step: Agree an information-handling policy before staff use AI tools.

Review data handling, retention, security, supported functions, costs and failure behaviour against the intended use. Marketing claims alone are insufficient. Consider vendor changes and a practical exit route before depending on the service for critical operations.

Next step: Document the intended data and controls required from the supplier.

Retrieval-augmented generation supplies retrieved information to a model to help produce a contextual answer. It can improve relevance but does not guarantee correct retrieval or interpretation. Access control and source quality remain necessary, especially for private knowledge.

Next step: Test retrieval and answers separately using representative questions.

Apply the user's permissions to retrieval and responses so the model receives only authorised information. A prompt telling it not to reveal secrets is insufficient. Test cross-role access and avoid logging sensitive retrieved content unnecessarily.

Next step: Attempt restricted questions using a lower-privilege test account.

Prompt injection occurs when untrusted content attempts to redirect an AI system's behaviour. It may appear in retrieved pages or documents. Treat external text as data and constrain tools and permissions rather than relying entirely on instructions inside the prompt.

Next step: Review which external content the AI reads and which actions it can perform.

Test representative inputs, difficult cases and failure conditions against explicit criteria. Assess factual quality and operational consequences, not just fluent wording. Keep appropriate human approval for consequential actions until evidence and safeguards justify the proposed autonomy.

Next step: Define evaluation cases and a safe fallback before enabling actions.

Check factual accuracy, tone, privacy and whether the promised action actually occurred. Use verified records and avoid unsupported performance claims. Approval requirements should match the message's risk, with correction and escalation routes for errors.

Next step: Test a generated update against the underlying completed-work records.

AI can assist with bounded tasks where permissions, evidence and recovery are well defined. Publishing, spending and changing live systems can have consequential effects and need suitable oversight. Producing a task list is not the same as completing the work.

Next step: Separate drafting, approval and execution and verify each final outcome.

can separate available model checks, observed AI referral visits and self-reported enquiries in reporting. These measures answer different questions and should not be blended into a guaranteed ranking. Use the results to guide evidence-based next work.

Next step: Review mentions, sources and actual enquiries separately with clear testing dates.