The checklist is a Zaher framework, not an industry standard. It applies to any provider, Zaher included, and every question asks for something you can see in writing or test yourself.
Why a sales demo tells you little
A demo usually shows one answer from one assistant, often in English. Three published findings explain why that is not enough to judge a provider:
- Answers change from run to run. SparkToro and Gumshoe ran 12 brand-recommendation prompts 60 to 100 times each on ChatGPT, Claude and Google's AI in November and December 2025. On ChatGPT and Google's AI, the chance of getting the same brand list twice was under 1 in 100; Claude repeated lists slightly more often (Search Engine Journal, 30 January 2026).
- The language of the question changes the sources. In more than 1,400 runs of 20 hotel questions in Arabic and English across five AI environments, 28% of the non-platform sources cited appeared in both languages (PRmoment, 26 August 2026).
- The answer itself is what people read. In browsing data from 900 US adults in March 2025, people clicked a traditional result on 8% of Google visits that showed an AI summary, against 15% without one, and clicked a link inside the summary on 1% (Pew Research Center, 22 July 2025).
So a credible provider measures many questions, many runs, several named engines and both languages, and connects what it finds to visits and sales.
The 9 questions to ask before you sign
This checklist is Zaher's framework, built from the research cited here and from Zaher's own audit work. Ask for every answer in writing.
| Question | Why it matters | What a good answer looks like |
|---|---|---|
| 1. Which engines are measured, by name? | Engines behave differently: in the same repeat-run test, Claude repeated brand lists more often than ChatGPT and Google's AI (SparkToro and Gumshoe, reported January 2026). | A written list of engines, and of the engines left out, so a result on one is never read as a result on all. |
| 2. Is Arabic measured as its own track, including dialect where customers use it? | Arabic and English answers drew on different sources: 28% of non-platform sources were shared (PRmoment, August 2026). | Arabic questions written the way customers ask them, reported separately from English, never machine-translated from an English list. |
| 3. How many times is each question run on each engine? | The chance of the same brand list coming back twice was under 1 in 100 on ChatGPT and Google's AI (SparkToro and Gumshoe, reported January 2026). | Several runs per question on every engine, and results reported as a share of runs, not as a rank in one answer. |
| 4. How many answers sit behind each number, and what change counts as real? | With answers this variable, a small sample moves by chance from month to month. | Every rate shown with its number of answers and its dates, and a stated threshold below which a change is reported as noise. |
| 5. Can I reproduce the evidence myself? | If you cannot re-run a finding, it is an opinion, not a measurement. | For every finding: the question text, the engine, the date, the full answer and the sources it cited. |
| 6. Is the baseline recorded before work starts, and are the same questions kept each quarter? | In Zaher's delivery experience, a baseline taken after changes go live absorbs the early gains, and a changed question set makes before-and-after comparisons meaningless. | A dated baseline before any change, a fixed question set for the term, and new questions tracked in a separate set. |
| 7. Is the result measured in my own analytics and sales data? | AI referrals were 1.08% of website traffic across 13,770 domains (industry benchmarks, updated 6 July 2026). Visibility and revenue both need measuring, in your systems. | AI assistants set up as their own channel in your GA4 and store data, compared against a pre-work baseline, with a like-for-like check where tracking changed. |
| 8. Does the provider promise a ranking or a position? | Google says no third-party tool has access to its internal ranking or AI systems (Google Search Central, updated 10 July 2026), and answers vary from run to run. | No promised position. A committed method, a fix plan and a re-measurement against your baseline instead. |
| 9. Who owns the questions, answers and reports? | A baseline loses its value if you cannot keep it when a contract ends or a provider changes. | A full export of questions, answers, sources and reports, in a format you can open without the provider. |
Red flags in a proposal
- Screenshots as proof. One answer, captured once, shows what one run said on one day.
- A single score with no engines, dates or sample size. You cannot tell what moved it, or check it.
- Arabic results built from translated English questions. Customers in Cairo, Riyadh and Dubai phrase questions in their own words, and the sources behind Arabic answers differ.
- Guaranteed positions. "Number one in ChatGPT" is not something any provider controls.
- Files sold as a Google lever. Google's own guide says Google Search does not use llms.txt files or special AI markup, and that creating them will neither help nor harm Search visibility (Google Search Central, updated 10 July 2026). They may still serve other services, but they are not a reason to buy.
Do you need a platform or an agency?
It depends on what your team lacks. If you have people who can write content, fix pages and work with publishers, a platform that measures regularly and tells you what to fix may be enough. If you do not, you need someone who also does the work: the content, the Arabic third-party sources and the technical fixes.
Either way, keep measurement and delivery connected. An agency that delivers without independent measurement grades its own work. A platform that measures without a fix plan leaves you with a report and no decision. Ask any provider who will re-measure, on which questions, and on what date.
How Zaher answers these questions
We hold ourselves to the same list. What we can state publicly today:
- Engines and runs (questions 1 and 3). The Foundation plan runs 30 questions, 3 times each, on ChatGPT, Gemini and Google AI Overviews, in 2 markets, for $49.99 a month. Enterprise: contact sales; 60 questions scaling up to 100; up to 8 AI engines. See pricing.
- Arabic (question 2). In our delivery work, Arabic is measured as its own track and never assumed from English. GEO Analysis compares results by language and market.
- Evidence (question 5). The full answer and the cited sources are kept for every run. How Zaher measures explains the method.
- Revenue (question 7). For Al Fouad Pharmacies, an Egyptian pharmacy chain, AI-referred revenue grew 11.9 times in the client's own GA4 between two consecutive 90-day windows (22 December 2025 to 21 March 2026, and 23 March to 20 June 2026). For a like-for-like read on a later window: chatgpt.com, the one AI source tagged the same way in both windows: revenue +667%, 5 May to 2 Aug 2026 vs the previous 90 days (GA4). Part of the growth reflects AI referrals being labelled more clearly (GA4 added a built-in AI Assistant channel on 13 May 2026, inside the second headline window), and the results come from the period of the work, not a controlled experiment. Details are in the case study, and the Analytics Suite covers AI traffic in analytics.
- Positions (question 8). Zaher commits to the method, the fix plan and re-measurement against your baseline, never to a ranking or a position in any assistant.
- Scale. 294,247 AI query runs, 6,365 audits and 5,854 domains measured, to 27 October 2026 (Zaher platform data).
For questions 4, 6 and 9, ask us to put the answer in the proposal, as you should with any provider.
Zaher is the Attention Layer for Arabic-first brands: it measures and improves what AI assistants say about a brand before a customer reaches its website. To run the same checks by hand first, use our guide to checking your AI visibility, and see our research for sector studies.