Product changelog

What has changed in the engine and the dashboard, with the month it went live.

Only changes you notice while using the product are listed. Internal work that changes nothing for the user does not appear here.

October 2026

• Searches for your brand, next to your AI presence. AI rarely sends the visit itself. People who read about you in ChatGPT usually search for you on Google next, and Analytics files that visit as organic or direct: the first contact appears in no report. What does leave a trace is more people searching for your name. Under Traffic there is a new card that sets, period by period, your brand searches in Search Console next to your AI presence over the same period, with its confidence interval and without the questions that already name your brand. With Search Console connected the series is weekly; if you upload exports, each one is a point and they are compared by daily average. Searches for other people with your surname and searches that look like bots do not count. And the card always says the same thing, even when both lines go up together: this is correlation, not attribution. A campaign or press coverage moves searches for your name just like AI does. With fewer than three points it gives no reading.

  

• What Google says when people search for you. Google has started showing its AI summary (AI Overview) when someone searches for a brand. In the US it already appears on most of those searches, according to studies published in September, but all of them measure the US and English. We measured it on Google Spain on 7 October. For the bare brand name the summary appeared for 4 of 10 brands in our portfolio, and in another test a few hours earlier, for 1 of 16. With "reviews", "pricing" or "what is" after the name it appeared in 25 of 30 searches. Those are exactly the searches made by someone who already knows you and checks you out before calling. In Sentiment & Authority there is a new card: once a month we search for your brand in those four ways, store what Google writes and its sources, and flag four things. First, when it states a reputation ("reviews are very positive", "4.9 out of 5") without citing any review platform. Second, when it spends part of the answer on another company with a similar name. Third, when it cites a domain with your name that you have not declared as yours. And last, when something contradicts the facts you confirmed in the Accuracy Score. In the portfolio, 3 of 10 brands had a reputation with no cited reviews, 2 a namesake and 2 a domain with their name to check. Counts are shown, not percentages: it is four searches. And it does not add to your Share of Answer. If you have Search Console connected, the searches come from the ones people actually make with your name, not from a template. And the card comes with a PDF plan with one action for each flagged item. Its goal is not to get Google to show a summary, because that cannot be requested. It is that, when Google shows one, it says the right thing.

  

• How you appear, not just whether you appear. Showing up when someone asks "which are the best companies for X?" and not showing up when they ask "which one has the best reviews?" or "which would you pick for a company like mine?" is visibility without positioning, and an average hides it. Under Competition there is now a table that splits your presence by the criterion the asker decides on (price, reviews, alternatives, fit for their case, specialisation, speed, proximity or plain "the best"), shows the reason AI gives for naming you in the sentence you appear in, how often it names you for the very reason asked (they ask about price and it names you for price) and which competitor wins each criterion, and with what argument. It runs on answers already measured, so it comes with your full history from day one. Questions that name your brand are excluded, every percentage carries its confidence interval, and below 10 answers we give no verdict. The full table, one row per answer with the query, criterion, intent, your status, competitors, sentence, attributes, sources and sentiment, downloads as "Positioning matrix" or connects to Looker Studio.
  And what to do about it. Below the table comes a plan, in order, with one action per criterion where the recommendation slips away or where you are named for something else: what to publish (prices with a figure, reviews and case studies, honest comparisons, who you are for, turnaround times, locations, proof of track record), who does it and how long it takes, a real question where the problem shows, and the third-party sites AI reads on that criterion that do not name you. It downloads as a PDF to hand to whoever will do it. It does not promise how much you will rise: it says what to do and where, and the following weeks of measurement will tell whether it worked.

  

• Can an AI agent complete the task on your site? More and more people do not visit websites: they ask an agent to search, compare and choose. AI Readiness has a new test: an AI browser tries to find your price, opening hours and contact route on your site, three times each, and reports which page it found them on, with the exact sentence, or where it got stuck. It separates what depends on your site (a bot check, a price that only lives in a PDF, a page with no text) from the noise of the agent itself, which also fails on its own. Read-only: it never fills in or submits any form. It does not count towards the score. We ran it on 12 websites in our portfolio: the agent reached the contact route on almost all of them and found the price on one in ten. See finding 11.

  

• Operational facts in the Accuracy Score. An agent acting for the user does not care about your founding year: it looks at price, opening hours, area, lead time and phone number. If AI gets one of those wrong, the agent calls the wrong place or rules you out over a price that is not yours, which is worse than not appearing. The Accuracy Score now puts those facts in their own subtotal, with their sample, and warns you if you are not tracking any yet.

  

• Delegated questions. The questions suggested when you add a domain, when you write those for a vertical and in the free audit now include a new intent: the user who does not ask for information but hands over the decision ("find me three, compare them and tell me which one to call"). It is how people who use an agent ask. No question is changed for domains already being measured: it would break their series.
• What is known about optimizing for agents (AVO), and what is not. The term is going around without sources. We have put it in order on one page, with what the data says (agent traffic, in-chat checkout only in the US, why agents fail on some sites and not others) and what cannot be measured yet.

September 2026

• What your customers ask you, turned into questions to measure. Until now, questions came from Bing, Search Console and sector patterns. The source every serious method recommends was missing: what your customers ask you before they buy. In «Discover prompts» you paste emails, your contact form, call notes or chat, and each question comes out as the one that person would ask ChatGPT: "Do you do succession plans?" becomes "Which consultancy helps a family business prepare its succession?". Questions naming your brand, after-sales questions and any line with an email or a phone number are dropped automatically, and we do not store the text you paste, only the questions. It is also the only source bots do not pollute: in our own Search Console, 160 queries were prompts from other tools (finding 09). The full method is in a four-step GEO method.

  

• Authority Share: of the citations to third-party sites, how many go to pages about you. In the cited-sources table, a domain with 601 citations did not tell you whether any of those pages was yours. Now every domain carries a second number next to it: the citations that go to pages dedicated to you (your name in the URL) or to pages that name you in their text, such as lists, rankings or news. To know whether a page names you we open it and look for your brand in the article, not in the menu or the sidebars, and we keep the sentence where you appear. When you expand a domain, every URL carries its tag: about you, names you, doesn't name you or not checked. The overall share comes with its coverage: what we could not open counts neither for nor against you. Your own social profiles and your competitors' sites are left out.

  

  And whether it pays off. Leaving out prompts that already contain your brand, we compare how often AI names you when it cites at least one third-party page about you with how often it names you when it cites third parties that don't. It comes with its confidence interval, and we only say there is a difference when the intervals don't touch. It is a correlation and it says so: AI may decide to name you and then look for a page about you.

  

• From citation to contact: does AI cite you on pages that sell? Crossing the Bing and Search Console reports with our own measurements turned up something uncomfortable: AI mostly cites guides, glossaries and comparisons, not the pages where people hire you. On joseredondo.es, our founder's site, only 2.6% of 3,665 Copilot citations in three months landed on a sales page. And the client with the most presence in our portfolio is precisely the one whose citations land on service + city pages. The new view shows what share of your citations lands on pages where someone can hire you; your presence in buying questions ("who should I hire for…?", "how much does… cost?") versus the rest, with its margin of error; which already-cited pages should send the reader to a sales page, and whether they do; and the sales pages nobody cites, alongside the buying searches you already get on Google. We only say you appear more or less in buying questions when the margins of error do not overlap; when they do, we say that too.

  

  Plus a contact log. You record each new contact and how they found you. ChatGPT often leaves no trace in Analytics, so without that question a customer who came through AI is counted as direct traffic and nobody attributes it. We do not estimate calls: a contact only counts once it is logged. What goes with it is adding "How did you hear about us?", with a "ChatGPT or another AI" option, to your form, your phone script and your booking page.

  

• Your LinkedIn, separated from the rest of LinkedIn. In the cited sources view, LinkedIn citations are broken down by kind of page (posts, articles, company page, profiles) and by engine, and your brand's own LinkedIn pages that AI cites are shown separately, with the question that brought them up. If your name has namesakes, you tell us which profile is yours and we stop guessing. These citations are not added to your Share of Answer. Which part of LinkedIn AI cites.
• Search Console no longer counts as demand what was probably made by a robot. Two kinds of robot never click: rank-tracking tools, which leave impressions far down the page, and AI visibility tools, which run their prompts against Google AI Mode ("as the director of…") and show up near the top. We now download Search Console filtered to your country, with the daily series and the country breakdown, and also use the report you upload; those rows are no longer used to suggest questions, for quick wins or to prioritise pages; and the dashboard warns when part of your impressions looks more like a crawler than a person, with the average position recalculated without them. What came out on our two websites.
• The technical score changes method: whatever stops you being cited no longer costs points, it blocks. If your page does not respond, carries noindex or nosnippet, blocks in robots.txt the crawlers AI searches with, or its robots.txt returns a server error, the audit puts it right at the top, in red, even if everything else scores 95. Blocking training crawlers (GPTBot, ClaudeBot, CCBot) no longer counts against you: it is your decision and it does not stop you being cited. The top level is now called "Well prepared" instead of "Leader", and from there the audit tells you that improving the site moves little. llms.txt weighs half what it did, because no search engine has confirmed it reads it. Audits run before 10 September are still shown, marked as the previous method, because they are not comparable with the new ones.
• The technical audit no longer passes a page that asks not to be used. We ran our AI Readiness audit and Cloudflare's agent readiness scanner over the same websites. We scored them between 93 and 98; Cloudflare, between 20 and 27. Most of that gap comes from measuring different layers, but reviewing ours turned up real faults. The engine did not read noindex or nosnippet, and Google documents that nosnippet keeps a page out of AI Overviews and AI Mode, so a page like that could walk away with a 95. It is now a failure that lowers the score. The sitemap was only looked for at /sitemap.xml; the Sitemap: lines in robots.txt are now read, and a sitemap index is no longer counted as if each sitemap were a page. Three new readings are shown without scoring, because no AI search engine publishes the threshold needed to turn them into a grade: how long your server takes to start responding, whether your robots.txt slows Bing down with Crawl-delay, and whether your server answers differently to ChatGPT's search crawler and to OpenAI's training crawler. That last one never states that you block OpenAI, because from outside there is no way to know, and it tells the two apart: rejecting the training crawler does not stop you being cited. Agent signals go from 10 to 12 and are also looked for at the older paths many sites still use; they still do not score. And audits saved in June and July stop showing a citation multiplier we withdrew in July for having no source. What came out when we measured 23 client websites.
• We no longer just tell you what AI gets wrong about you: we tell you how to fix it. The Accuracy Score and the sentiment sources diagnosed well and stopped there — and a diagnosis without a treatment is not a deliverable, it is a reason to complain. Now every error comes with its plan: what to do, why, and in what order. The order is not an opinion: it comes from how many errors each action fixes and how much of somebody else's permission it needs, so what you can do today on your own always sits at the top. And the plan downloads as a PDF, with a cover and with its limits printed inside, so it can be sent as it is. The first of those limits: the sources listed under an error are the ones AI cited in that same answer, not proof that the wrong fact lives inside them. It is an ordered lead, not an accusation — before writing to anyone, open the page and check. The page-refresh plan and the local visibility plan download the same way.

  

• Every source tells you whether there is a door, and which one. This came from something very concrete: a partner downloaded the 315 sources of his niche, scored them by hand from 0 to 10 and sent the file back with a note — "without a justification, the score is an arbitrary number that helps you decide nothing". He was right, and his scale was in fact a taxonomy of who controls the door, which means it can be computed. Now every domain in your opportunities carries its score with the criterion attached to the number: a 9 is a channel of yours or a listing you fill in yourself; a 5 is somebody else's site with a person to write to; a 1 is a global platform with no door at all. These are two axes and they are never multiplied: how much a source is worth, and how hard it is to get in. What it surfaced the moment it was switched on: across three different portfolios, the top ten opportunities included support.google.com, google.com and ibm.com. They now rank lower — but they are never hidden, because a hard door is not the same as no door.

  

• No more suggesting you publish on your competitor's website. A user spotted it looking at his own dashboard, and he was right. The cause was silly: competitors were recognised by name and not by domain, so "Profound" did not match tryprofound.com. Measured across the whole portfolio, 118 competitor domains were being proposed as targets across 26 brands, several of them in the top slots. They are now recognised by the domain already on file, the rivals the engine itself detects in the answers are added, and you can mark any domain as competition yourself with one button — in case you know one we do not. It undoes just as fast.
• Upload your sitemap and we tell you which pages to start with. Until now the dashboard only knew the pages AI already cites, which are precisely the ones that need no work. The sitemap closes that blind spot and costs nothing, because you publish it yourself. Crossed with your Search Console report, out come the pages that already have demand on Google and zero AI citations: that is where the work pays off soonest. It takes the XML, a sitemap index — it follows the children — or a list of addresses pasted by hand. If you have not uploaded the search-engine report yet, no priority is invented: you are told what is missing to compute one. And the last-modified date is served as declared, not verified, because many CMSs set it to today on every page.

  

• Classic Google Search, inside the dashboard and kept apart from AI. The same Search Console ZIP carries the ordinary report — clicks, impressions, CTR and position — and until now it was discarded. It is now shown, in its own block and never added to the AI one: they are two different surfaces of the same search engine, and a number mixing them means nothing. It is not there to measure AI visibility; it is there to have in front of you the floor you are starting from when deciding which page to touch — a query with impressions and no clicks already has demand, it just needs a push.

  

• Add a line of business and the questions write themselves. Extending a vertical used to mean typing twelve questions by hand, one per line. There is now a button per vertical: it proposes the questions split by buying intent — use case, problem, comparison — and respects your plan's quota. It never overwrites what you have already written: it appends. And if we do not know which city you work in, it invents no local questions; measuring a geography that is not yours for months is worse than not measuring it.
• Copilot, the fourth engine measured. It joins ChatGPT, Perplexity and Google AI Mode. It runs monthly rather than weekly, and the reason is arithmetic rather than commercial: at its price per query, measuring it every week does not fit the measurement budget without cutting into the other three.
• Upload your own Bing and Search Console reports, and they are analysed for you. Bing publishes which of your pages Copilot actually cites, and Search Console how often you appear in Google's AI answers. Neither is an estimate from anyone: the search engines sign it themselves. Neither exposes it via API yet — we checked, and Microsoft has it "in the backlog" — so the only route is the CSV their dashboards export. Until now you had to email us that file for us to import it by hand; now you drop it into the dashboard and that is it. It works out on its own which search engine each file comes from, takes the Search Console ZIP exactly as downloaded, and the three separate downloads of the Bing report complete each other instead of overwriting. Upload the wrong file and one button undoes it: the Search Console export without the AI filter has exactly the same columns as the right one, so no program can tell them apart — better to say so than pretend otherwise.

  

• Competitors you did not know you had — and adding them costs nothing. The dashboard reads the answers already measured for you and surfaces the brands in your sector that AI names and you are not tracking. These are not guesses: every name has been verified to appear literally in one of your answers, and you are told in how many. Adding one recomputes your entire history instantly and at no cost: because we store the full text of every answer, there is no need to ask AI again — we re-read the ones already paid for. In a real portfolio, a newly added brand showed up across 215 measurements from the previous two months. A warning we also give inside the dashboard: adding competitors lowers your share of mentions and may lower your rank, because the sector now has one more brand. You did not get worse: you were comparing yourself against fewer players than there are, and the new number is the right one.

  

• The dashboard tells you where to start. The action plan grew to over a hundred tasks sorted by a score, and a sorted list of a hundred things is still a list of a hundred things. It now shows three, and the rest is grouped by kind of work — content you are missing, places you should be, pages to refresh, pages competing with each other — because each group gets done in one sitting. The score stops being a number without a scale and says "high impact" instead. And the summary opens with a sentence rather than four acronyms: "AI mentions you in 3.7% of your sector's questions; you are seventh of twelve". The margin of error is still right below, which is where it belongs.

  

• How often you are cited as a source, by vertical and by engine. That column existed and was never computed: it read zero for everyone, which reads as "nobody cites you". It now counts the real citations to your domain in each segment. And where the data does not exist — clients imported from another tool do not carry it — we say so instead of printing zeros.

  

• Sentiment: what you are accused of, and in which answer. When an answer speaks badly of you, the dashboard shows the actual sentence, the terms that flagged it and a reputational-risk warning, with a link to the full answer. It used to show the beginning of the captured text, which in Google AI Mode is the page's navigation bar.
• Request a page straight from the finding. Where the dashboard spots a question you do not appear in, you can ask us to write it. It is logged with the question and with who is beating you there today.
• Accuracy Score: showing up is no longer enough. It measures what share of the times AI states a fact about you it gets right. It only counts when the model actually makes a claim —not mentioning your founding year is not a miss— and the percentage always comes with the number of claims evaluated and its interval. It runs over answers already measured, so it appears with your full history the day you switch it on, without measuring anything new.

  

• Sources and sentiment: which pages feed the bad answers. It crosses the tone of each answer with the sources that answer cites, and flags which domains show up mostly when AI speaks worst about you. It is correlation, not causation, and it says so: everything is compared against your own average and sources with insufficient sample are not listed.

  

• Your data inside your ChatGPT or Claude (MCP). Connect the account to your assistant and ask about your visibility in plain language, without opening the dashboard. Every query is checked against your own scope: asking for another brand's id is cut off before anything is read.
• Direct export to Looker Studio and Google Sheets. Four tables —weekly trend, prompt by prompt, sources and accuracy— behind a single URL that Sheets reads with IMPORTDATA. The metric lands in the report where GA4 and Search Console already live, instead of sitting apart.
• Agent readiness in the audit. Ten new checks on whether your site is ready for AI agents (negotiated Markdown, Content-Signal, MCP server card, API catalog…). They do not score, on purpose: almost no site has them and no engine has confirmed using them. It is a dated technical inventory, not a failing grade.
• Public niche indexes. AI visibility rankings for one specific sector, with the same questions for every brand and the sample size on show.
• vilnux Research: what comes out of measuring. An open section publishing what shows up while measuring — Google AI Mode's citation collapse in August, how often an engine changes its answer to the same question, how few sources two engines share — each with its sample, its method and what the data does NOT allow us to claim. First-party figures can be re-derived from the corpus: 19,026 measured answers across 22 brands.

August 2026

• Annual billing on all four plans. Twelve months for the price of ten, and checkout from the homepage now takes one click instead of two.
• Content delivery is now a single document. It used to mean opening three files per piece across a folder tree; each weekly delivery now arrives as one Word file with its own index.
• One more engine, and cadence by plan. Measurement is scheduled according to what each plan includes, instead of applying the same frequency to everyone.
• Exactly what you get in each plan. Dashboard sections compared one by one, with no small print.

July 2026

• Email is asked for after the result. You see your audit first; the address is only needed if you want to keep it.
• Immediate cancellation. Cancelling stops the service and its cost the same day, with no waiting for the cycle to end and nobody to email.
• Payment from outside Spain. All four plans can be bought from any country.
• Cost per client in view. For agencies, what it costs to measure each domain in their portfolio.
• Engine outage warning. The system checks whether the engines we sell can actually be measured today, rather than finding out on Monday.

June 2026

• Confidence intervals across the board. Every percentage is published with its margin and the number of observations behind it. It was the change that moved the numbers the most — and the one that made them defensible.
• Three repetitions per question, by default. A single pass cannot tell a real gain from a coincidence. (Since August 2026 the weekly pass is one run per engine and the figure is computed over every week in the period.)
• Method stamp on every measurement. Which engine, which model, and with or without web search is now recorded, so a series cannot average two different things without saying so.

Frequently asked questions