// Decision guide
Measure AI visibility as a repeatable set of observations, not one universal ranking: keep a fixed prompt set, record brand mentions and citations separately on each AI surface, verify crawl and index eligibility, use Google Search Console and Bing Webmaster Tools where data is available, and track self-reported leads. A single ChatGPT answer is a sample, not a score. This guide explains the method Tarek Abou Rjeily uses for AI-visibility audits in Lebanon and the Gulf, based on platform documentation checked on August 17, 2026.
Build a prompt set around real buyer intents, markets, and languages, then keep the wording fixed between measurement rounds. Run each prompt in a fresh or temporary conversation with web search enabled and record the surface, mode or model, country, date, and whether the run was valid. Personalization, location, model updates, and changing web indexes can all alter the answer, so the context belongs beside every result.
Record brand mention, owned-domain citation, and verified third-party citation as separate fields. A brand can be named without evidence; a directory profile can be cited while the owned domain is absent; and an owned page can be retrieved for background without appearing as a visible citation. Those are different problems and require different fixes.
| Weak measurement | Useful measurement | |
|---|---|---|
| Ranking | One claimed universal 'AI rank' | Mention and citation rates by fixed prompt, surface, market, language, and date |
| Evidence | A screenshot of one favorable answer | A saved result log with valid runs, citations, competitors, and the unchanged prompt set |
| Technical check | robots.txt looks fine in the repository | Live status, headers, rendered HTML, canonical, sitemap, crawler access, and engine index reports |
| Attribution | Assume every direct visit came from AI | Referrer and UTM data plus optional self-reported discovery on qualified leads |
| Impact claim | Visibility rose after a change, therefore the change caused it | Annotate the change, allow for recrawl, compare trends, and state that the result is observational |
A page cannot earn a search-grounded citation if the relevant system cannot discover or index it. Check the live HTTP response, robots policy, noindex and snippet controls, canonical URL, sitemap inclusion, internal links, and whether important text exists in the rendered HTML. Also check the CDN or firewall: allowing a bot name in robots.txt does not help if its verified requests are blocked upstream.
Search and training controls are separate decisions. OpenAI documents OAI-SearchBot for ChatGPT search and GPTBot for potential model training. Anthropic separately documents Claude-SearchBot, Claude-User, and ClaudeBot. Perplexity documents PerplexityBot for search. A visibility audit should preserve access for the search-time crawlers the business wants while treating training access as an explicit policy choice.
Google says its existing search fundamentals apply to AI Overviews and AI Mode: useful people-first content, crawlable links, clear textual information, good page experience, and structured data that matches the visible page. It also says there is no special AI schema, no required chunk size, and no need to rewrite pages purely for generative systems.
For a professional service, the strongest content is usually non-commodity evidence: named work, the exact problem and delivered workflow, who performed the work, limitations, dates, inspectable sources, and client-approved outcomes. Consistent external profiles and genuine client-authored evidence help resolve the entity. Fabricated reviews, thin location pages, and inauthentic mentions damage trust rather than create it.
Google announced dedicated Generative AI performance reports in Search Console in June 2026, initially for a subset of sites. They report impressions and participating pages for generative features in Search and Discover. If the property has access, use that report alongside normal index coverage and URL inspection; if it does not, do not substitute a third-party estimate and call it Google data.
Bing Webmaster Tools' AI Performance report shows citation activity across Microsoft Copilot, Bing AI summaries, and selected partner experiences. Its page-level citations, grouped grounding queries, topics, intents, and citation share are useful trend data, but Bing explicitly warns that citations are not rankings, authority, traffic, or proof that one content change caused a movement.
AI answers often produce dark traffic with no reliable referrer. Keep UTM and referrer-host capture, but add an optional self-reported question such as 'How did you find us?' with ChatGPT, Perplexity, Google AI, Claude, and Copilot options. A cited page matters more when it produces a qualified conversation.
Run the same measurement set on a fixed cadence, annotate deployments and indexing requests, and compare trends after discovery has had time to catch up. Report what changed and what was observed, not a guaranteed causal claim. Model and index changes can move results even when the site did not change.
It can use a clearly defined internal scorecard, but there is no universal AI rank. A defensible score names the prompt set, surfaces, countries, languages, date range, and weighting, while retaining the underlying mention and citation data.
A mention means the answer names the brand. A citation means the answer visibly references a source URL. Track owned-domain and verified third-party citations separately because they show different levels of source control and entity corroboration.
Google states that it does not use llms.txt for Google Search and that the file neither helps nor harms Google Search visibility. It can still be maintained as an optional summary for other tools, but it should not replace crawlable, useful HTML pages or index monitoring.
Monthly is a practical operating cadence for a small service business, with an additional check after material content changes have been crawled. Keep prompts stable and avoid daily interpretation of noisy answer changes unless monitoring a specific incident.
At minimum: the dated prompt and result log, technical eligibility findings, index coverage evidence, entity and external-proof gaps, page-level priorities, a scored backlog with owners, and a re-measurement date. It should distinguish observed facts from recommendations and never guarantee citations.
Product facts can change. These primary sources were checked on the guide's update date; verify plan details again before buying.
The first step is a short, free discovery call — including an honest build-vs-buy recommendation. Message me on WhatsApp or email tarekabourjeily@outlook.com.