Market intelligence for the layer where AI decides.

Weekly and automatic across every engine and surface. Not an estimate: the real answer each engine gives.

• Two routes per engineTraining memory (what the AI knows about you without the internet) and live web search (what it answers when it searches). The contrast between the two tells you whether the problem is the brand entity or presence on the web.
• Buyer questionsThe questions a buyer actually asks, put to the engines in the voice of a persona (CEO, CMO, technical decision maker) and by decision stage. The persona is a framing we write and send ahead of the question. The same question also runs with no persona at all, to measure how much the persona changes the answer.

“Knowing the market” is not a slogan: it is a list of verifiable things, and each one changes what the numbers say.

• Measured where the buyer isIn the markets where collection is localised, the questions run on that country's surfaces: what a buyer there sees, not what an API returns to a data centre in another country. In Portugal it has worked like this from day one.
• Buyer questions, not translationsWritten the way people write in your market. In Portugal, “qual o melhor ferro de engomar até 50 euros” (which is the best iron under 50 euros). A translated list measures a market that does not exist.
• Local competition, by nameThe system tells an agency from a software tool, a comparison site from a shop, a direct peer from an adjacent player, and share of voice only counts whoever competes for the same contract.
• Findings that only exist up closeWhoever decides the recommendations made to a market's buyers is often outside that market. In Portugal, we found product pages still waiting to be translated and buyers sent to the Spanish Amazon. We measure this, with a date.

And the first published study on how AI recommends products in Portugal is under way, with its protocol made public before collection began.

Almost every AI visibility measurement counts mentions. A mention can be your name in the middle of a list, and it can be the sentence “and the best of these is X”. Counting both the same way erases the one difference that matters to whoever is selling.

• Citation ratePresent, or not. One answer, one yes or no.
• Share of voiceThe mentions in the category. Being one of three is worth more than one of twelve.
• Share of recommendationThe choices in the category. Only answers where the assistant picks somebody count.

The denominator for choice is the answers that pick somebody. An answer that explains, lists and compares without picking has no pie to divide, and stays out. All three metrics use the same scale on purpose: two numbers read side by side have to be comparable.

Three real answers, and what each one is worth

These are answers about destaque.ai, the company behind Periscopy, measured by Periscopy itself. We use our own because the ruler we apply to others has to apply to us. The excerpts are translated; the originals are in Portuguese and unedited.

The criterion is not invented per client. It has been fixed in writing since the 616-answer study that produced it: a place in an ordering counts as a choice, so does a verdict on a criterion naming one brand, or the closing sentence. Belonging to the set (“the big three”) does not count, nor does a niche superlative. When in doubt, it does not count.

The distinction is no longer only ours. Marketeer, a Portuguese marketing publication, ran the study on 6 September 2026 using the same pair of words, under the headline “EDP is the Portuguese brand AI chooses most”.

• Named and classifiedEvery name the AI mentions is classified: direct competitor, tool, or media. We do not count an SEO tool as a rival of yours.
• A dossier eachWhich engines cite it, the exact sources feeding those citations, the questions where it appears and you do not, real excerpts of how the AI describes it, and the share of voice trend, week by week.
• What it is doingServices it announces, who it sells to, content themes, channels where it is active (blog, LinkedIn, podcasts, press, reviews) and recent moves. Facts with a source, nothing else.
• Perception mapEvery brand placed by how much it appears and by how it is described: leaders, niche, controversial, laggards.

A technical director and a finance director do not ask the same thing about the same product, and AI does not answer them the same way. Asking “what is the best X platform” without saying who is asking measures an average that describes no real buyer.

The concept: a synthetic persona

A synthetic persona is a buyer written out in full and used as a variable in the measurement. It goes in front of the question, so the assistant answers as it would answer that person, rather than nobody in particular.

It is not a slide persona. A marketing persona lives on a deck and changes nothing; a synthetic persona is executable: it changes the question that goes out, and therefore the answer that gets measured. Two personas on the same product produce two different measurements, and the gap between them is the information.

• Role and segmentWhether it is B2B or B2C, and which side of the table they sit on.
• ContextThe sentence that goes in front of the question.
• GoalsFeeds the generic and feature questions.
• PainsFeeds the comparison and objection ones.
• Decision criteriaWhere the choice is played out.
• VocabularyThe words THEY use, which are not the brand catalogue's.

The last field is the one most people forget. A brand describes itself in its own vocabulary; the buyer searches in theirs. Where the two do not meet, the brand does not show up, and it is not for lack of content.

An incomplete persona is a label, and the product says so. Every profile shows how much of it is filled in. A persona with only a name changes nothing in the measurement, and presenting it as though it did would be selling a reading that does not exist.

The limit, written down because it is honest: a synthetic persona is a hypothesis about a buyer, not the buyer. It does not replace talking to customers. What it does is make the measurement conditional and comparable: instead of an average that describes nobody, a result per profile, which can be argued with and corrected because it is all written down.

When somebody asks “which GEO consultancy do you recommend in Lisbon”, the assistant does not go looking for that sentence. It breaks it into the questions it thinks it needs to answer first, searches each one, and only then writes. That middle step is the most actionable thing in this measurement, and almost no tool shows it.

A real case. Asked “How much does a GEO audit cost?” in a destaque.ai audit in the week of 7 September 2026, Gemini ran six searches before answering. Here they are as they came out, none rewritten, in Portuguese because the question was:

• auditoria GEO preco
• "auditoria geotecnica" preco (other industry)
• quanto custa uma auditoria GEO marketing
• quanto custa "auditoria de barragem" (other industry)
• preco auditoria GEO
• GEO Generative Engine Optimization preco auditoria

Two of the six searches went to geotechnics and dams. A third of the search budget spent in another industry, and the fix is positioning: attaching the term to the right category.

Three things you only see here

• The words it searched withIf the question was about price and it searched for pricing table and how much per month, the page missing from your site has a name.
• In how many sub-questions it found youA question can be split into four. Appearing in one of the four is not the same as not appearing, and you know which one.
• Who is in your place in the othersFor every sub-question where you are absent, the domains it read instead of yours. That is the list of pages to write, in order.

The engine searches for brands by name before it knows what it will find. On another question that same week, about proven case studies, Gemini ran nine searches and six of them named a specific brand. That is the candidate list it already carries. Being on it is a different measure from being cited, and it comes first: what is never searched for cannot be found, and no work on the page fixes that.

Markets slip. On the same pricing question, ChatGPT ran a single search and it ended in “preço Brasil”. On a Portuguese question, that pulls competitors from another market into the answer your buyer reads.

Honest note: not every assistant declares this step. Where it is not declared, Periscopy says it did not see it, rather than counting zero.

Week after week, the internal searches either repeat or they do not. The two mean opposite things, and both are actionable. If it converges on a handful of sources and terms, being in those terms is a finite job with an end in sight. If it brings vocabulary from outside your sector, the category has no owner yet: that is a window, and windows close.

Four measures a week, each pointing at a different action:

• Overlap between searchesHow much the assistant repeats itself week to week. High, the job has an end in sight; low, the category has no owner yet.
• Searched for your nameThe brand is already an entity it recognises and goes looking for. What is never searched for cannot be found.
• Read your siteYour pages reach the crawl. That is the floor, and it is not guaranteed.
• Read it and did not cite itIt got there and the page did not convince. The most expensive problem, and the one that most easily goes unnoticed.

The last measure is the one that changes conversations. “We do not show up” and “we show up, it read the page, and it picked someone else” call for opposite work. The first is presence; the second is the page failing to convince, which is the most expensive problem and the easiest to hide.

The thresholds for what counts as much or little repetition do not live in the code: they live in the method, and they change when the measurement forces them to. What the product does is count; the reading is kept separate on purpose.

Beyond measuring where you are cited, Periscopy audits what the engines find when they reach your site: the foundation once, and then page by page, with no practical ceiling. A site of 2,000 pages is analysed in rounds that continue on their own.

• The foundationRobots, llms.txt, structured data, entity. Each signal opens and explains itself: what it is, whether it is yours to fix or depends on a third party, and the correction already written.
• Two scores per pageTechnical (can the crawler read it) and answering (does it answer what it promises), with the concrete problems and the fix for each.
• Patterns, not noiseWhat repeats across the site is aggregated. One duplicated title across 500 pages is a template problem, fixed once.
• Ready for agentsThe emerging standards, markdown for agents, Content Signals and discovery, measured as readiness and kept out of the score: we tell you what is coming without inflating what counts today.
• Products agents can readIf you sell online there is one more layer. A shopping agent reads price, stock, delivery time and returns from your data, not from the page. We measure the fill rate across a sample spread over the whole catalogue, then product by product. In a catalogue of 10,000 items what counts is not the one lucky page, it is the percentage, which is where Merchant Center fails you.

Not everything is taken. Part of the questions and sources in your category have no owner yet. Periscopy finds that territory and tells you when it changes.

• Unowned questionsWhere the AI still recommends nobody. Whoever arrives first with quotable content takes the answer. Buying-intent questions are flagged as priority.
• Unowned sourcesThe domains the engines cite in the category where no competitor dominates: the cheapest way in.
• Positioning to claimThe angle nobody holds in the answers, grounded in the week's data.
• AlertsWhen a competitor enters a question that was yours, when a new name appears in buying questions, or when an open question finds an owner.

A supermarket that wants its discount mentioned, a consultancy that wants to be named when someone asks about price: that does not belong to one question, it belongs to a subject. You declare it once and it holds for every question on that subject. The following week the measurement says whether it started being said, with the sentence that proves it.

The starting point is frozen. It is read the day you declare it, not the day you go and look at the result. Otherwise the before moves with the after and the movement disappears.

With no public proof, writing it is the first step. We never ask an engine to assert what is not published. With the piece in place, the work is getting the engine to it; without it, the work is a different one, and the plan says which.

A claim nobody has read yet shows as unmeasured. Not as zero. They are different things and they lead to different work.

• A brief per questionEvery open question has an attack button: it generates the brief of the content that can win it, with title, structure, and the sources where to plant presence.
• Eight dimensions, four horizonsEvery gap tied to a concrete action, with severity and effort estimated. Appearing is not just a website problem, so the plan is not just technical.
• Before and afterEach completed action is dated, and the panel shows citation before and after it. Results measured, not promised.

Eight dimensions

• Technical
• Content
• Entity
• Authority
• Social
• Authority on site
• Measurement
• Positioning

Four horizons

• Horizon 1, weeks 1 to 2
• Horizon 2, weeks 3 to 6
• Horizon 3, weeks 7 to 12
• Horizon 4, 90 days and beyond

• Every answer, keptEach answer from each engine, stored by week, with citation, position, tone and the competitors detected in that answer.
• Next to your trafficSearch Console, AI bot visits to your site, and a correlation panel between visibility and traffic, side by side.
• ReportingMonthly PDF report and a weekly email digest, with the highlights and one action for the week. CSV export of everything.

Once connected, “what does my brand have to do this week” is answered from the report, not made up. Four tools, all read-only: the assistant reads what we measure and changes nothing.

How to connect. In Claude’s connectors, as a remote MCP server, with the address https://periscopy.ai/api/mcp. Sign in with your Periscopy account, no key to paste, and each connection sees one brand only.