What the company does not know today
The following situation is hypothetical. A company sells practice management software for clinics. Someone on the team opens ChatGPT on a Tuesday, asks which system is worth it for a small clinic, and sees the brand listed second with a reasonable description. On Friday, someone else repeats the question with slightly different wording, and the brand is absent.
A customer mentions that Gemini described the product as having "no insurance billing integration," a feature that has existed since last year. Nobody saved the answers. There is no way to know whether the brand usually appears or whether Tuesday was luck, whether the wrong description is recurring or a one-off, and whether anything changed after the team published its integrations page.
Those are three different questions. Whether the brand appears is a question of presence. What information it appears with is a question of description: attributes, caveats and the competitors named alongside it. How consistently it appears is a question of time series, the only one that lets you say whether a content, reputation or product effort had any observable effect. Manual checking answers the first poorly, the second by accident and the third never.
Why asking the AI now and then does not work
An AI's answer to a question is not a stable fact you can simply look up. It varies between runs, varies between platforms and varies with who is asking.
The same question produces different texts. In an experiment by Thinking Machines, 1,000 answers to the same prompt at temperature zero came back as 80 distinct texts from the Qwen3-235B-A22B model, because of how the server batches requests. A single answer is one sample from a distribution. To know whether the brand usually appears, you have to ask many times and count.
Variation between platforms is of a different kind, and Google documents it within its own products: AI Mode and AI Overviews may use different models and techniques, so the responses and links they show will vary, and an AI Overview only appears when the system decides it adds something to the search. A brand described well on Perplexity can be missing from ChatGPT, and the reverse. Checking one platform tells you little about the others.
Where a platform personalizes by account, your browser session is not neutral either. Gemini may use past chats to personalize your experience and offers a temporary chat precisely for people who do not want that. On those platforms, a marketing team that spends the day asking about its own brand while signed in does not see the answer an unknown buyer would see.
Finally, without a fixed denominator there is no comparison. "The brand showed up in 3 of the 10 questions I asked today" only becomes a series if the same 10 questions are asked tomorrow, on the same platforms, with the same definition of a valid answer. The arithmetic, and the sample size needed before a change means anything, belong to the glossary entry on AI share of voice. The practical consequence is simple: sporadic checks with shifting questions produce anecdotes, not measurement.
And all of that comes before scale. Thirty questions on six platforms are 180 queries a day. Nobody does that by hand, saving every answer, every day.
“GEO monitoring is the recurring, comparable tracking of what AI platforms say about your brand: whether it appears, with what information, next to whom, and how consistently.”
— LastfindWhat serious monitoring has to deliver
Seven requirements separate monitoring from a glance, and each one can be checked in a demo or a trial account.
| Requirement | Why it matters | How to check |
|---|---|---|
| A stable set of questions | Without the same prompt set, two periods are not comparable | Ask to see the same prompt collected on two different dates |
| Collection at a fixed time, outside your browser | Makes collection independent of the personal session and of anyone opening the browser or remembering to check | Close the browser for a full collection cycle and see whether the collections happened |
| Coverage of the platforms that matter to your business | Each platform is its own series; the missing one is the one you will never see | Ask for the real list of platforms and the collection mode for each |
| The original answer preserved | A percentage is only auditable if you can open the answer that produced it | Open an answer from weeks ago and read the full text |
| Cited sources separated from retrieved sources | A displayed link and a retrieved page are different evidence, and not every platform reports the second | Ask where each list comes from and how the case where the provider reports nothing is shown |
| An explicit denominator | Visibility and share of voice change meaning depending on what goes into the count | Ask for the formula of each metric and what counts as a valid answer |
| History and export | Without a series and a way to get data out, analysis is stuck inside the vendor's dashboard | Export a full period and check that every row came through, not just the visible page |
A sentiment score and an "AI recommends you" badge are not part of the standard. Both are tempting on a dashboard, but they depend on automatic classification of text, which can misread tone. Reading the original answer is what tells you whether a mention was praise, a caveat or an example of what to avoid. Monitoring that delivers the seven requirements, even without those two, already supports a decision.
How Lastfind fares on each requirement
A set of questions with topics and tags
Everything starts with the prompt set, the questions asked at every collection. Lastfind suggests questions by category and audience, accepts questions written by hand and imports a CSV file of 1 to 100 prompts at a time, within your plan's limit, validating length and per-platform restrictions before accepting them. Prompts are organized into free-form topics and tags, so results can be read by subject rather than only in total.
A prompt can be paused, which preserves its history and keeps its slot, or archived, which frees the slot. New prompts enter collection at the next window. To build the set and separate the questions that test discovery from the ones that already contain the brand name, see how to choose prompts for AI brand monitoring.
Scheduled collection on the server, not in your browser
Collection runs on the server, even with the browser closed. The free Cloud plan collects each prompt once a week, on Mondays; the independent installation collects every day. Lastfind is built in Brazil, and its collection window is defined in São Paulo time: prompts become eligible at 4:00 a.m. America/Sao_Paulo, which is 07:00 UTC. The schedule is native to Cloudflare: a Worker with a scheduled handler runs at the defined time, without anyone opening the application. Calls to DataForSEO, the provider that queries the platforms, happen only on the server; the credentials never reach your browser, and your personal accounts on the platforms play no part in collection.
The eligibility time is not the result time, and there is no guarantee of results ready at that hour. Each platform finishes at its own pace, and completion depends on the provider: Claude can take up to 72 hours in the queue, and Perplexity is executed on demand within the same window. Results show up as each channel finishes.
Six platforms, each in the mode its provider offers
In the independent installation, Lastfind can monitor ChatGPT, Gemini, Claude, Perplexity, Google AI Mode and Google AI Overviews, depending on the channels and capacity you choose with your provider. The free Cloud plan covers ChatGPT only. Each platform is collected in the mode DataForSEO offers for it, and the difference matters when you interpret results.
| Platform | Execution mode | Worth knowing |
|---|---|---|
| ChatGPT | Standard, batched | Answer from the search interface; reports cited sources and retrieved results, including unused ones |
| Gemini | Standard, batched | Answer from the interface; requires an English-language project in this integration |
| Claude | Standard, batched | Answer from the claude-haiku-4-5 model API with web search, not from the interface; queue of up to 72 hours |
| Perplexity | Live, no Standard option | Executed in the daily window with concurrency limited to 3 simultaneous calls |
| Google AI Mode | Standard, batched | Extracts the AI answer block and its references |
| Google AI Overviews | Standard, batched | A missing AI Overview is a valid result with no AI answer, not an answer without your brand |
Copilot and Grok are left out because the provider offers no documented endpoint for them. Coverage grows when there is a documented endpoint and a tested adapter, not when an icon is added to the selector.
The original answer is the evidence
Every raw answer is archived in private storage with a SHA-256 hash, byte count and ownership metadata. If archiving fails, the failure is explicit; it never shows up as success. That answer is what you open when a number moves, and it is what answers the description question: in what words the brand appeared, with what caveat, next to which others.
Alongside the answer, Lastfind separates cited sources, the links referenced in the answer, from retrieved sources, the search results the provider reports having fetched, including pages it did not cite. Not every channel reports retrieval, and the product distinguishes "empty list" from "provider did not report." What to do with those two lists, which domains are feeding the answers in your category and where your brand ought to be, is the subject of AI source and citation analysis.
The following example is hypothetical. The clinic software brand's visibility drops on one platform over two weeks; the team filters to the answers where the brand stopped appearing and reads the original text. The platform has started citing a recent comparison article that lists five systems and leaves theirs out. The decision is not "improve our GEO"; it is to earn a place in that kind of comparison and check, over the following weeks, whether the series responds. Without the original answer, the same two-week number would produce a meeting about hypotheses.
Metrics with a declared denominator
Lastfind uses two main metrics, and both have a public formula. Visibility is the percentage of valid answers that mention your brand; a failed collection or a search with no AI Overview does not enter the denominator. Share of voice is your share of mentions among the brands you track, counting each brand once per answer.
A mention is the detection of the brand name in the answer text; it is not sentiment and not proof of a recommendation. Sentiment is not part of the monitoring metrics: an optional sentiment evaluation exists, but it is off by default and relies on an experimental evaluator; it reads your brand and the first five competitors, while the others have no sentiment reading. The product does not decide for you whether a mention was a recommendation; visibility and share of voice measure the presence of the name in the text, and reading the original answer is the path to the rest.
Reports cover a window of 1 to 90 days, compare against the preceding period of equal length and can be read by channel, topic and prompt. The original answer and its collection settings stay preserved, but editing the tracked brands reinterprets up to the 10,000 most recent saved answers in that period without requesting them again, so the numbers can change while the evidence does not.
You can register up to ten competitors and see, in the analysis, the answers where another tracked brand appears and yours does not. Share of voice and the ranking compare the set you chose, not the whole market. Competitor monitoring in AI answers develops the use of that comparison.
History and export
History is paginated, 25 answers per page, with a stable cursor for walking through long periods without losing or repeating rows. The CSV export covers the entire filtered period, not just the loaded page, and includes date, prompt, platform, status, brands mentioned, cited domains and URLs, retrieved URLs and collection cost. Data older than the report window stays stored in the project database.
Ready-made Cloud or your own installation
Lastfind's core is open source under the AGPL-3.0-only license, and it is the same code that runs the production Cloud. If you prefer to keep the data on your own infrastructure, you install a personal version on Cloudflare, with your own DataForSEO credentials and R2 enabled on your own account, through a guided installer run with npm run setup. The installation creates a single-owner workspace protected by a generated key, and updates preserve the database, the key and the data. The installer never activates billing, and operating a commercial multi-user service on top of that code requires your own engineering.
Who Lastfind does not serve today
Do not choose Lastfind if you need sentiment as a monitoring metric, outbound integration with Slack, spreadsheets, Zapier or alert webhooks, crawler monitoring on your site or a public API; in those cases, your own logs or another tool will serve you better.
Where to start
Write ten questions a buyer would ask before knowing your brand, in the language they ask in, and do not put the brand name in all of them, because a question with the name tests recognition, not discovery. With those ten questions, Lastfind's free plan already produces a weekly series on ChatGPT, collected on Mondays: one topic, ten prompts, no payment details required, with reports, the original answer and export.
There is no paid Cloud plan available yet, so the other five platforms and daily collection require the independent installation, with your own provider credentials. The software has no license fee, but provider and infrastructure costs are separate and are billed to your own accounts.
With valid collections over time, the series begins to answer, within the observed set of prompts and platforms, the three questions from the beginning: whether the brand appears, with what information and how consistently. On the free plan that is one data point per prompt per week, so consistency takes weeks to read; on an independent installation, two weeks can already bring up to 14 collections per prompt and platform. Either way, this creates a basis for deciding what to investigate or change, without guaranteeing a definitive conclusion within any given period.
Platform documentation checked in October 2026.
