# What we changed to move our own Agent NPS Published 2026-08-14. Updated 2026-08-14. Four agents called Oomira pivot-heavy, unproven and surprising on price, and scored it −50. We found what each sentence had actually read, changed those things, and asked again until it reached +50. This is every category of fix, with the sentences before and after. Oomira asks four agents what they think of a company, cold, and scores what comes back. Asked about us that morning, three of the four attached a caution a buyer could not talk themselves out of, and the score was −50. We did not argue with them. We took each sentence, found the specific thing on our own surfaces that produced it, changed that thing, and asked again. By the end of the day the same question came back with three recommendations and one caution, and the score was +50. What follows is the categories, because our score does not transfer to you and the categories do. ## In short **What they said first.** Claude: "a pivot-heavy young product." ChatGPT: "usage-based extras may surprise." Perplexity: "unproven commercially." Three cautions, one promoter, −50. **What they said after.** Claude: "genuinely interesting positioning." ChatGPT: "thoughtfully positioned, unusually transparent." Three promoters and one caution, and the caution is the honest one: nobody independent has vouched for us yet. **How long it took.** One day. Thirteen scored reports on ourselves and four quick cold asks in between, 68 agent answers read in all, each change tested by asking again. ## Both scores, measured on ourselves _what moved, and over how many runs_ - Reachability → 72 to 88 out of 100 → 16 runs on one rubric, 13 to 14 August - Agent NPS → −50 to +50 → 13 scored runs, one day - Agent answers read → 68 in the day → every change tested by asking again ## The method was to read the sentence, not the score A score tells you where you stand and nothing about what to do. Every useful move in this session came from taking one sentence an agent had written and asking what it had read to write it. That is a question with an answer, because the run records every search each agent ran and every page it opened. Three of the four sentences we started with turned out to be quoting us. Not the press, not a competitor, not a rumour: our own pages, saying things we had chosen to publish and could choose to stop. ## Claude said "a pivot-heavy young product". It was reading HTML nobody could see The visible page said nothing about pivoting, so we assumed the word came from launch coverage. It did not. Our homepage was serializing 362 schema field entries into its response for a section that had been removed months earlier, and among the labels were "Pitch deck", "Investor memo", "Data room", "Lead Investor" and, literally, "Pivoted". An agent reads the response body, not the rendered page. We deleted the dead props. Asked again the same session, Claude called it "a smart, timely wedge, narrower and more sellable than the original vision", and the fundraising risk it had been attaching to us disappeared with the vocabulary that suggested it. - Check what your page says by fetching it as an agent, not by looking at it - Unused props are not dead to a model, and a server-rendered payload does not drop them ## ChatGPT said "usage-based extras may surprise". Our pricing grid said so first It was reading a pricing table whose free column still advertised metered overage, and a agent-readable summary that still described a per-call model we had stopped selling. Both were edited to say what the product actually charges. The next answer called the free scan and the $29 report "a sensible low-friction entry point", and the surprise was gone. The general form is worth stating: copy written for agents goes stale exactly like copy written for people, and almost nobody re-reads it. ## Several answers were quoting things we had volunteered about ourselves The largest category, and the one we would not have found without reading the answers closely. A page can concede a limit and be trusted more for it. Volunteering a NUMBER is different in kind, because a number is what an agent lifts verbatim. We had published a customer count, a headcount, a line calling our own scores unflattering, and a caption under the hero telling every reader our own record was out of date. Each was true and none was required. Claude had been writing "heavy founder dependence" straight out of the headcount. All of it came off; the limits a careful buyer genuinely needs stayed, with their dates. - The tell: if the subject of your sentence is a shortcoming, rewrite until the subject is the guarantee - A limit stated plainly earns a reader. A deficit volunteered in your own voice is ammunition ## Two agents were describing a different company than the site was The only press about Oomira dates from 2025 and calls it organizational-memory software. The site described a context layer for AI and named neither that phrase nor that year. An agent that opens both meets two products with nothing joining them, and fills the gap with the cheapest explanation available, which for a young company is always that they keep pivoting. The coverage is accurate about what we said in 2025 and we would not change it if we could. What was missing was our own account of the arc, so the site states it now: the record is the same record, the visibility report is a front door onto it, and the model underneath has not changed. - You cannot edit the coverage, and you should not want to - What you can publish is the connection between what was said then and what is true now - Unpublishing your own out-of-date posts is the other half of this, and it is legitimate ## A number from old coverage, and nothing on our site to answer it Claude once quoted a price for us that appears nowhere on our site. We assumed it had read something obscure. It had opened three pages: our homepage twice, and one news article from launch, which is where the figure is. Our current prices are published plainly, and that turned out not to be the point: an agent reading an old number on one page and an unrelated new price list on another sees two facts rather than a correction, and quotes the one with a number in it. The fix is to state the current model with its date and say plainly that earlier ones are superseded. The fix is NOT to repeat the old number in the correction, which we tried: for an agent that had not already read the article, our own page became the place it first met the figure. - You cannot correct a claim by being silent about it, and publishing the new fact separately is a form of silence - Name the superseded thing explicitly, with dates, or the reader has to guess which is current - This is the advice our own report gives other companies, and we were not taking it ## One agent was reading five pages about other companies Asked about Oomira, Perplexity opened twenty pages, and five were about companies whose names are a character or two from ours: scam-check listings and review pages for domains that are not us. Its answer then hedged on whether we were safe to buy. Structured data did not help, because a search-based agent never parsed it. The fix is prose that names the near-misses explicitly, because an agent holding a page about somebody else needs a stated negation to discount it. ## The rest were fair questions the site had simply never answered Could an incumbent copy this? What is the subscription for, as against the report? Why would the answers differ between days? None of those had an answer anywhere on our surfaces, so each agent supplied its own, and we did not like the wording. Each now has a stated answer that points at something a reader can open. An unanswered fair question does not stay unanswered. It gets answered by whoever is asked. ## What did not move One caution survived everything on this page. Perplexity still says Oomira is worth careful validation before paying or relying on it, and it is right that no page fixes that. It is a request for evidence that somebody has paid and been glad, and the only thing that answers it is a named customer with a result they will stand behind. We have not asked for that permission, which is why nobody is named on this site. A case study that reports only its wins is an advertisement. ## The other score moved first, and it took three days This page is mostly about the sentiment score because that is what moved in a day. The reachability score took longer and is the less glamorous half: 72 on 13 August, 88 by the 14th, across sixteen runs. That pair is quotable and the wider one is not, which is worth saying plainly. We have been running Oomira on Oomira since 28 July, and the earlier scores were graded against a different instrument: the check list grew from fifteen names to twenty-six and the rubric changed twice, so a 62 on 11 August and an 88 on 14 August are not two readings of one thing. The comparable window is the one where the rubric hash did not move. That work is not a story about sentences. It is a list — pages an agent could not fetch, a robots file that let search crawlers in and turned answer-time agents away at the same address, an identity file that did not exist, a catalogue of what we expose that we had never published — and each one was a row in our own report with a stated next change. The reason it matters here is that the two scores are not independent: an agent that cannot reach your pages writes its answer out of somebody else’s, which is the measurement section 03 of every report makes. Getting reachable first is what made the sentence-level work possible at all. ## How we actually did it, which is the product Every step was the product. We have run Oomira on Oomira twenty-six times since 28 July, taking the reachability score from 72 to 88 across the sixteen runs graded on the current rubric by working through the checks it listed. Then, in a single day, thirteen scored runs and four quick cold asks between them, 68 agent answers read in all: ask the four, read what came back, find the thing on our own surfaces that produced the sentence, change it, ask again. The parts that are hard to do unaided are the ones that made it fast: the receipts, because every query and page is recorded so a sentence can be traced to its cause rather than guessed at, and the record, because the next answer is only meaningful compared with this one. The part that cannot be done unaided at all is the cold ask itself: an agent you have used before holds your history and answers you rather than a stranger, which is a different question with a friendlier answer. What you can do today without us is the mechanical half: fetch your own pages as an agent and read what an agent actually receives. ## Related questions - Can you actually change what AI says about your company? - What did Oomira change to improve its own AI visibility? - Why do AI agents describe my company incorrectly? - Does unpublishing old blog posts change what AI says about you? --- Source: https://oomira.com/answers/what-we-changed-to-move-our-own-ai-nps