Practical answers about source quality, citations and recommendations. Updated September 7, 2026.
One missing mention does not identify the cause. Check whether search was used, what the question requested, which sources were cited and whether the business fits the stated need. Then inspect source access, entity ambiguity, missing offer details and contradictory public information. Repeat comparable observations before deciding which page needs work.
Inspect the audit process →No. A provider can commit to defined research, source improvements, implementation and reporting. The assistant controls the generated answer. Keep advertising placements and organic recommendations separate, and require the measurement protocol behind any claimed placement or uplift.
Review the agreed scope →Start with the existing search foundation. Google says its established SEO practices remain relevant to AI search experiences. Examine whether your page answers the full buyer decision and whether the generated answer uses it accurately. A search position, an AI citation and a recommendation describe different observations.
Google's AI optimization guide →A new site can publish accurate, original information, but useful content alone does not guarantee discovery or selection. Explain the actual offer, make the sources accessible and establish relevant public evidence. Start with decisions you can answer well instead of producing many interchangeable pages.
Build a focused answer hub →Choose the surfaces your buyers actually use and document their separate behavior. Consumer ChatGPT, a provider API, Perplexity search and Google's AI experiences are not interchangeable tests. Record surface, model when visible, search mode, language, market and time so a change in setup does not masquerade as progress.
Use the measurement method →It adds a disciplined way to connect buyer questions, public source quality and observed AI answers. Research a decision, improve the page that owns the answer, substantiate its claims, record the resulting responses and revisit the gaps. SEO, editorial review and product clarity remain part of the same work.
Follow the connected resource library →Check the response status, robots rules, index controls, canonical URL, meaningful HTML content and any firewall or login challenge. Allowing a crawler is one access condition, not proof that the page was indexed or cited. OpenAI documents OAI-SearchBot for search separately from GPTBot for training.
OpenAI crawler documentation →Structured data can describe supported facts on a page, but it must match the visible content. Google states that no special schema is required for its AI search features. Choose relevant markup for the page and validate it; do not treat the presence of FAQ schema as a recommendation score.
Read Google's requirements →Treat it as an optional file for tools that explicitly support it. Google says it does not use llms.txt for Search. A maintained sitemap, crawlable links, appropriate indexing controls and useful visible content should take priority. The existence of this file is not evidence of AI visibility.
Check the documented role of llms.txt →Use sources relevant to the claim: an official registry for a regulated status, a partner page for a real partnership, documentation for a product capability or a dated customer study for a measured result. Keep names and scope consistent. A count of mentions or links cannot establish their accuracy or commercial relevance.
Review evidence and ranking signals →Preserve the question, full answer and sources. Correct the factual problem on pages you control and request corrections on relevant third-party sources when appropriate. Retest after the source is accessible and updated. Retrieval and answer generation vary, so a corrected page does not guarantee the next response is correct.
Run the next improvement cycle →There is no universal timeline. Deployment, crawling, indexing, source selection and response generation are separate events. Set a baseline window and a review cadence, record what changed, and report the observations available at each review. Do not convert an implementation schedule into a promised date for recommendations.
Plan the implementation →Resolve a blocking access or index issue before expanding content. Then choose a commercially important decision where your source is incomplete or inaccurate and where you can supply real evidence. Record the expected observable change, its owner and its test. Prioritization should follow the finding, not a generic points checklist.
Prioritize a defensible fix →No universal conversion uplift should be assumed. Measure qualified inquiries and confirmed sales in your own analytics or CRM, keep the attribution window explicit, and allow unknown origin. Referral traffic, self-reported discovery and a saved AI recommendation provide different evidence and should not be added into one inflated total.
Build a comparable benchmark →Use a fixed set of questions and save each complete response with its date and setup. Report brand mentions, linked source citations and explicit commercial recommendations separately. Include valid non-mentions; exclude failed requests from the observed-answer denominator while displaying missing coverage. Keep branded identity questions outside non-branded recommendation rates.
Document the observation panel →Arrow GEO separates its technical website checklist from observations returned by configured provider APIs. An unavailable provider is unmeasured. A provider API observation is not a consumer search position. Use the product page to check access and limits, and the free audit to define the source improvements and observation scope you need.
Inspect Arrow GEO and its limits →Inspect the website readiness studies, review the France DPE implementation case, or choose another decision in the buyer question library. The full resource map connects the steps.
Arrow GEO
Review your public source pages, inspect the gaps, and decide what to improve with Arrow.
