# Measure AI recommendations with saved answers, stable prompts and explicit counting rules

> Source: <https://dev.to/newtation/measure-ai-recommendations-with-saved-answers-stable-prompts-and-explicit-counting-rules-1b3d>
> Published: 2026-10-22 04:30:47+00:00

Schema, entity markup, llms.txt and crawler access are implementation questions. Whether an assistant recommends your business is an observation question. I would settle the observation rules before using technical changes as evidence of progress.

I’m Ansh Khandelwal, Co-Founder & Strategist of Newtation, a GEO/AEO consultancy that helps brands get cited by ChatGPT, Perplexity and Gemini. The [Newtation Authority Engine](https://newtation.ai/authority-engine/) focuses on making existing credibility legible and strengthening authority around relevant buyer questions.

Newtation helps brands get cited and recommended by AI systems through the Newtation Authority Engine, which focuses on making existing credibility legible and strengthening authority around relevant buyer questions.

My measurement rule follows from that focus: keep the answer behind the number. A recommendation report should show what a buyer asked, what the assistant answered and why the response received its label.

A missing recommendation does not tell you which implementation task to choose. Start by checking whether the relevant evidence is accessible.

For Google AI Overviews and AI Mode, supporting pages must be indexed and eligible for a search snippet. Google also says no special AI schema or extra machine-readable file is required for these features. That is the boundary I would apply when someone proposes schema or llms.txt as the next measurement milestone. Completing an implementation task does not establish that a buyer received a recommendation. [Google’s technical guidance](https://developers.google.com/search/docs/appearance/ai-features).

Keep entity markup work tied to the facts being represented. If an answer describes the business incorrectly, save the exact error and identify the authoritative correction on the website or important profiles you control. The next observation should check whether the description is accurate.

I would keep an access issue, a description error and an absent recommendation as separate work items. They require different evidence to assess whether the work helped.

Use separate fields for these outcomes:

| Outcome | Counting rule | 
|---|---|
| Mention | The business name appears in the answer. | 
| Recommendation | The answer affirmatively presents the business as suitable for the buyer’s question. | 
| Citation | The answer visibly links to a source. | 
| Accurate description | The answer gets the offering and buyer fit right. | 

A negative mention remains a mention, but it should not receive recommendation credit. A citation to your website does not automatically mean the answer proposes your business as a provider. A third-party citation about your company belongs in a different column from a citation to your own domain.

Write these rules before collection begins. Otherwise, reviewers can quietly change the meaning of a successful result from week to week.

Tool definitions deserve the same attention. Ahrefs counts a brand once per answer even if the name appears repeatedly. It also distinguishes cited pages from pages found without an inline citation. A report using different rules needs an explicit explanation before anyone compares the totals. [Ahrefs metric definitions](https://help.ahrefs.com/en/articles/15501968-ai-visibility-metrics).

A saved response without its conditions is an incomplete observation. Record the assistant, product surface, selected mode, date, language, location settings and whether web search was used. Preserve the exact prompt and the full answer, including visible sources.

A collection record can use a simple JSON structure. This is an illustrative record shape, not a test result or a required standard:

```
{
  "prompt_id": "",
  "exact_prompt": "",
  "platform": "",
  "product_surface": "",
  "selected_mode": "",
  "collected_at": "",
  "language": "",
  "location_settings": "",
  "web_search_used": null,
  "collection_status": "",
  "answer_archive": "",
  "named_brands": [],
  "brand_mentioned": null,
  "business_recommended": null,
  "owned_site_cited": null,
  "cited_urls": [],
  "description_accuracy": "",
  "reviewer_note": ""
}
```

The empty values matter: this example supplies no invented observations. Your collection process fills the fields with actual records.

Use a fresh conversation for every core question. Keep follow-up tests separate and retain the preceding conversation. Perplexity documents model and search-mode choices, along with context retained during follow-ups. Those settings belong beside the answer rather than in someone’s memory. [Perplexity Pro Search documentation](https://www.perplexity.ai/help-center/en/articles/10352903-what-is-pro-search).

Build prompts from sales calls, enquiry forms, support conversations and customer interviews. Preserve the buyer’s language.

For a small team, I would cover finding providers, comparing options and checking suitability. Keep branded questions separate. Asking whether a named firm is suitable measures something different from asking which firms deserve consideration.

The Dubai formation examples show why context matters. “Who can help me set up a company in a Dubai free zone?” asks for providers. “What should I check before choosing a company formation adviser?” asks for decision criteria. Those suggested questions should not be treated as interchangeable observations or as test results.

Keep markets and languages distinct too. An English prompt about Dubai does not stand for an Arabic prompt about the same service.

Freeze a core set for reporting. Put new questions into an exploration group first. When a question joins the core set, log the change and establish a new baseline. Otherwise, a changed score can reflect a changed test.

Use separate reporting rows for ChatGPT, Perplexity, Gemini, Google AI Overviews and Google AI Mode.

Google says AI Overviews and AI Mode may use different models and techniques. An AI Overview also does not appear for every search. Record “no AI Overview” separately from an Overview that appeared without your business. [Google’s explanation of AI features](https://developers.google.com/search/docs/appearance/ai-features).

A timeout, blocked request or failed collection is another distinct outcome. It does not establish brand absence. Keep failures outside the valid-answer denominator and report them separately.

Repeat the core checks on a fixed weekly schedule. When an important answer changes, use additional repeat runs. Preserve every valid response rather than selecting the screenshot that makes the report look strongest.

Recommendation coverage is:

```
valid answers recommending the business / all valid answers collected
```

Show both counts beside the percentage. Label the result “recommendation coverage in our tracked sample.” The sample does not represent every conversation on an assistant.

Owned-site citation coverage is the proportion of valid answers with a visible citation to your website. Count an answer once even when it cites several of your pages. Keep third-party citations about the business separately.

Description accuracy needs agreed review labels. Accurate, partly accurate, wrong and insufficient detail give the reviewer a practical starting point. I would prioritise a wrong service description over a small change in recommendation order because the description can mislead a buyer about fit.

Qualified demand belongs alongside these measures. Track identifiable assistant referrals, enquiries and opportunities. Preserve a buyer’s discovery response as self-reported evidence instead of assigning the whole sale to AI.

Citation share requires a stated denominator. Microsoft defines Citation Share for a particular grounding query and calls it observational, rather than a ranking or traffic-share measure. Its reporting scope covers Microsoft Copilot, Bing and select partner experiences. [Microsoft’s reporting announcement](https://blogs.bing.com/search/2026/6/New-AI-Visibility-Insights-in-Bing-Webmaster-Tools-Intents-Topics-Citation-Share-Compare/).

Estimated impressions need a similar boundary. Ahrefs bases its metric on Google search volumes associated with prompts where the brand appears. That is a demand proxy under its method, not a direct audience count. [Ahrefs metric definitions](https://help.ahrefs.com/en/articles/15501968-ai-visibility-metrics).

Before attributing a jump to your latest changes, check the measurement system. Ahrefs documents that additions to its prompt database can produce mention spikes. It advises checking whether other brands rose together and whether the chart carries a platform change note. [Ahrefs guidance on sudden spikes](https://help.ahrefs.com/en/articles/16249245-reasons-for-sudden-spikes-in-brand-mentions).

For BUSINESS SETUP CONSULTANCY, UAE, the published case records a move from unranked to a primary recommendation after 90 days. The recorded result was 83 AI citations, 41% AI citation share, 101.6K impressions, +79% reach growth. The engagement included 80+ citation-ready assets. [Published case studies](https://newtation.ai/case-studies/).

The buyer question was “cheapest way to set up a company in Dubai free zone”. That makes the recommendation commercially specific.

The supplied summary does not disclose the complete prompt set, platform mix, rerun protocol or metric formulas. I would ask for those definitions before comparing this result with another dashboard. The case is a published engagement result, not an independent experiment or a forecast for another business.

Give each issue an owner, a reason and a check that can show whether it was resolved. Fix access problems. Correct wrong business facts. Improve pages where a buyer constraint remains unanswered. Inspect the external sources supporting competitor recommendations and pursue inclusion only where the business belongs.

Keep a dated change log. Where practical, leave a comparable question group untouched while improving another. Treat that comparison as directional evidence, without claiming causation.

For the Newtation Authority Engine, the weekly decision is concrete: which buyer question lacks usable evidence, which description needs correction and which change needs another observation period.

*Written with AI assistance and checked against the sources linked inline. Newtation works on AI visibility: [newtation.ai](https://newtation.ai).*