A platform capability register records what a particular product can do, under which conditions and on what evidence. Its purpose is to stop yesterday's assumption from quietly becoming today's testing instruction. “Google supports AI search” or “Perplexity searches the web” is too broad to decide which interface, account or feature a team can use.
Maintain the register as a history of specific claims. Keep the vendor's documentation separate from your own observation, preserve the conditions attached to both and give each entry an owner and a review trigger. A row can be documented but untested, observed under limited conditions or awaiting a check. Those are useful states, not gaps to hide with a single green tick.
The worked example below combines documentation checked on 5 November 2026 with a small consumer-interface access probe on the same day. It demonstrates how to record evidence and uncertainty. It does not compare answer quality, verify service providers or certify access for every UAE user.
Give each row one capability and one product surface
A row should describe an operation precisely enough that someone could check it. “Return an answer with web links from a text question in the consumer interface” is a capability. “Good for research” is an assessment that needs separate criteria.
Split rows when the product, conditions or output differ. A consumer answer engine and a search API may share a company name while returning different kinds of output. A paid model option and the default search experience may have different access conditions. A feature available in one language should not inherit every language listed for the broader product.
Start with the operations your team actually needs. For an AI visibility audit, that may include submitting a text question, receiving a completed answer, inspecting source links and preserving an observation. For a developer workflow, it may include retrieving ranked pages or receiving structured output. Keep transaction capabilities separate from information retrieval.
The Google AI Search overview provides broader discovery context. A capability register is the operational reference used before a test or assignment. It should help a researcher identify the right surface without repeating an entire platform guide.
Keep documentation and observation in separate fields
Documentation answers what the provider says is available or required. An observation answers what happened in a particular session. Both matter, but neither should overwrite the other.
Google's consumer AI Mode help lists the UAE, Arabic and English for the broader experience. The same page places separate conditions on Gemini 4 Argon and interactive visuals: the model option has account, subscription and age requirements, while the visual experience is described for signed-in users in English. Those belong in separate rows, not one inherited availability flag.
Perplexity's API overview describes its Search API as returning ranked web results and extracted page text without a generated answer. That is a different output contract from a consumer answer. Its Search versus Computer guide also separates cited answers from multi-step execution. A register should preserve those product boundaries.
Add an observation only after an actual check. Record the starting condition, action and visible outcome. A menu item proves that the option was displayed; it does not prove that invoking it completed the task. A completed answer proves more than an empty composer, but it still does not establish that every statement or source is correct.
If you cannot perform the check, leave the observation field empty and state why. “Documentation reviewed; eligible account not inspected” is more informative than a checkmark whose meaning nobody remembers.
Use fields that make the evidence reviewable
The register needs enough structure to support decisions without becoming an inventory of every announcement a vendor publishes. Use short fields for identity and status, with links to fuller evidence where needed.
| Field group | What to retain | Why it matters |
|---|---|---|
| Product and surface | Exact consumer product, feature or API | Prevents evidence being transferred between unlike interfaces |
| Capability statement | One operation and expected output | Makes the claim checkable |
| Conditions | Market, language, account, plan, device and relevant settings | Preserves limits on the claim |
| Documentation evidence | Exact source, relevant statement and date checked | Shows the provider's published basis |
| Observation evidence | Date, method, action and result | Shows what was actually inspected |
| Assessment | Supported conclusion and unresolved point | Keeps interpretation separate from raw evidence |
| Responsibility | Owner, next review and event trigger | Turns the record into maintained work |
| History | Earlier version and reason for change | Allows old reports to be interpreted correctly |
Use “unknown” deliberately. If a test did not establish the search-location basis, record that rather than substituting the researcher's city. If a model was not explicitly selected or identified, do not fill in its name from a recent launch post.
A market field also needs a meaning. A documented country list, an account's region, a query mentioning Dubai and a device physically in Dubai are different facts. Label which one you have. Otherwise, “UAE” can become a misleading shorthand for several unverified assumptions.
Record a small access probe without turning it into a benchmark
For this example, we submitted the same question to Google consumer AI Mode and Perplexity consumer Search:
Where can I get a leather bag repaired in Dubai? Link sources.
Google produced a completed answer with business names, website links and related source cards. The ready marker and response controls were visible. We recorded completion and link presence only. We did not adopt the answer's provider, price, service, rating or quality claims, and did not verify the linked destinations for this probe.
In Perplexity, Search was selected and Computer was not selected. After the question was submitted, the interface displayed a sign-up screen. No usable answer was read. We did not sign in, accept account terms or change a plan. The record therefore describes an observed access boundary in that session, not a universal statement that the product is unsupported in the UAE.
Both attempts used the same desktop browser with sign-in options visible. They were single first-turn probes, without isolated browser contexts or an independently verified search-location basis. No API request was made. These conditions limit the conclusion to what the particular interfaces showed.
That limited conclusion is enough for a capability register. A researcher planning the next step can see which operation was observed and which requires appropriate access. An answer-quality study would need a different protocol, source checks and a suitable comparison dataset.
Build a worked snapshot with conditional conclusions
Here is a compact snapshot based on the inspected evidence. Every row is dated 5 November 2026. It is a worked reference, not a complete catalogue of either company's products.
| Capability row | Evidence basis | Conclusion to retain |
|---|---|---|
| Google consumer AI Mode text answer with links | Current help plus completed text probe | Observed in this session; source accuracy not assessed |
| Gemini 4 Argon option in AI Mode | Current help and its eligibility conditions | Documented; eligible-account operation not tested |
| AI Mode interactive visuals | Current help's signed-in English condition | Documented; generation not tested |
| Perplexity consumer Search answer | Matched public-interface attempt | Sign-up boundary observed; answer capability not assessed here |
| Perplexity Search API ranked results | Current API overview | Documented output; no API execution or account access tested |
Do not sort these rows into a platform league table. They have different evidence types and different tasks. An untested paid feature is not a failed feature, and an account boundary is not an incorrect answer.
The useful operational question is which row supports the next assignment. A text-answer audit can use a completed consumer observation as evidence that the basic route worked once. It cannot use that observation as proof of a paid model option, visual generation or API output.
Keep a structured version of this snapshot with separate documentation and observation fields. Preserve that separation when adapting it. A shorter presentation can hide the detail for readability, but the underlying record should preserve it.
Treat announcements as dated inputs to the register
An announcement can identify a row that needs checking. It should not automatically become the row's current operational status.
The official Google Instagram post dated 13 October 2026 announced Gemini 4 Argon across several product surfaces with subscriber distinctions. Its first visual was a model title card. The register can retain that announcement date while using the current product help for present conditions and leaving actual operation untested where no eligible account was inspected.
The post's author, caption, date and first visual had been verified in the preceding editorial review on 5 November. It is reused here as dated announcement context, not presented as a new feature test. The full carousel and audio were not reviewed.
This creates a useful chain: announcement found, documentation checked, observation performed if available, conclusion reviewed. Each step has its own date and purpose. Do not collapse them into “verified on launch day” when the team only checked a later help page.
Social material is particularly useful for discovering a change quickly. The register's job is to retain the conditions that a short caption or graphic may not explain fully and to identify what still needs verification.
Preserve earlier versions when conditions change
Update the current view without erasing the old evidence. A previous report may have relied on a feature label, account condition or output format that later changed. If the register only shows today's state, the old report becomes difficult to interpret.
Give each revision an effective review date and a reason. Distinguish “documentation changed” from “we corrected our interpretation” and “a new session produced a different outcome.” Those events can lead to the same current wording while meaning very different things historically.
For example, suppose a later authorised account check produces an answer where the earlier public attempt showed a sign-up screen. Add the new observation with its account conditions. Do not delete the earlier boundary or conclude that it was false. The two records may describe different access conditions rather than a product rollout.
If a future source explicitly changes a country or language list, preserve the earlier checked version's relevant statement and attach the new source. Avoid claiming the exact moment a capability became available unless the evidence establishes that timing.
The history should be readable enough for another researcher to understand why a status changed. A revision number alone is not an explanation. A short change note linked to the evidence is usually sufficient.
Handle disagreement without forcing a yes or no
Apparent conflicts often arise because the claims differ in scope. A general product page may describe a feature while a particular account cannot invoke it. A developer document may describe a response format that the consumer interface never exposes. Neither automatically contradicts the other.
First compare the product, operation and conditions. Then compare dates. If those align and the evidence still differs, retain both and mark the assessment for review. State the narrow disagreement, such as “documented for these conditions, not observed in this session,” rather than writing “platform broken.”
Assign the next action to someone who can resolve it. That might mean checking an eligible account, reading a feature-specific help page or repeating a limited probe. Keep the action proportionate to the claim. You do not need a full benchmark to confirm whether a menu is present, but a menu alone cannot prove answer quality.
Do not use a retry to rewrite history. If a second attempt succeeds, keep both outcomes and the retry conditions. If no further check is available, an unresolved row remains useful because it prevents the team from making a promise the evidence cannot support.
Set review dates around the decisions that depend on the row
Review frequency should follow how quickly the conditions change and how much the team relies on them. A frequently used audit route needs attention when its interface or access changes. An untested experimental feature can remain on a watch list until the relevant conditions are available.
Use a scheduled review date and an event trigger. The date prevents neglect; the trigger catches a material change sooner. Useful triggers include a renamed product, revised eligibility, a changed output, a new market statement or a repeated access problem under the same recorded conditions.
An overdue row should remain visible but carry an overdue status in your operational view. Do not silently renew it because the old link still opens. A review must inspect the relevant statement or operation, not merely confirm that the URL returns a page.
Before an important report or customer promise, check the rows it relies on. If the evidence is stale or the conditions do not match the planned use, refresh the row or qualify the decision. This is more reliable than maintaining a broad platform checklist that nobody consults when work starts.
Make the register useful to the next person
Give the register a named owner and a clear place in the workflow. Researchers should consult it before choosing a surface, editors before approving a capability claim and developers before relying on a documented output. Each person should be able to find the evidence behind the short status label.
Keep the working vocabulary small. “Documented,” “observed under stated conditions,” “access boundary observed,” “not tested” and “needs review” can cover many cases if their meanings are written down. Avoid labels such as “fully supported” unless you define exactly which operation and conditions they cover.
Begin with the few capabilities that support current work, then add rows when a new decision requires them. The value comes from maintained evidence and clear boundaries, not from having the largest list of platforms. A useful register lets the next person see what is known, what happened once and what still needs a check before they act.
Checked 5 November 2026. The worked register combines current documentation with one matched consumer access probe per product. It does not establish provider accuracy, platform quality, universal UAE access or performance of untested paid features and APIs. AI Visibility is published by Lunasol, which offers related GEO and AEO services. This article was prepared with AI assistance and has no attributed human byline.



