OperStack vs asking ChatGPT: a weekly record with page fixes

What do OperStack and manual ChatGPT checks do?

OperStack takes over a defined recurring task: 10 buyer questions every week in ChatGPT and Perplexity, a record of the businesses named and page edits to review for $35 monthly. Manual chats cost $0 on ChatGPT Free. Their value depends on how consistently you preserve questions, settings, dates and the resulting answers.

1. The recurring question set. A trend needs identical inputs. Watch asks ChatGPT and Perplexity one fixed set of ten questions every seven days with web search on, so a move from 4 of 10 to 6 of 10 can be investigated against the same question list. By hand, the wording drifts, the settings drift, and the week you were busy leaves a hole. SparkToro and Gumshoe’s research found that two answers to one prompt share the same brand list less than 1 time in 100, which is why a single chat cannot carry a decision.

2. Rivals by name, counted. A chat shows the names in one answer. Watch counts them across all ten questions, so you see which business takes your place most often, and it measures three rivals beside you each week. Even the free check sorts every name it finds into three groups: rivals, non-rivals like directories and law firms, and names it could not identify.

3. The reason and the repair. OpenAI says only that ChatGPT ranks search results using multiple factors. A recommendation chat does not systematically inspect your pages for a clear opening statement. OperStack scores ten of your pages out of 100 and writes the priority page fix with proposed text tied to a page finding and claims supported by your site.

4. No personalisation in the way. Your own ChatGPT may know your business from months of chats, and OpenAI says memory can shape answers and even the search query. Watch asks through OperStack’s code, not your account, so your memories and custom instructions play no part.

5. One email instead of a spreadsheet. Every seven days from the day you subscribe, one email says what changed. By our estimate, the same work by hand takes 1 to 2.5 hours a week.

6. Inside the rules. OpenAI says automatically extracting data from ChatGPT is prohibited under its Terms, so a custom extraction script should use the API rather than scrape the ChatGPT website. Watch is that work already done.

Start with the free check: three ChatGPT questions and up to ten pages scored, no signup. Watch costs $35 a month, or $299 for a year, on the pricing page; the first 7 days are free and you can cancel at any time.

Can I just ask ChatGPT if it recommends my business?

Yes, and it is a sensible first look. ChatGPT Free costs $0 and includes web search, and OpenAI says people who are not signed in can use web search too. What one chat cannot tell you is whether the answer you got is the usual one, or what to change if you were left out. Treat a single chat as one card drawn from a large deck: informative, but not a measurement.

What does asking ChatGPT yourself cost compared with OperStack?

A manual check can cost $0 because ChatGPT Free includes search. OperStack Watch costs $35 each month or $299 for a year, paying for the recurring checks, page scoring and written fixes. Optional ChatGPT plans and a custom API script are separate purchases with different limits and responsibilities.

The like-for-like starting point is ChatGPT Free at $0 against the OperStack free check at $0. A paid ChatGPT plan is optional for a manual look; the Watch subscription pays for the recurring record, comparisons and page work.

Put the time next to the price. At 1 to 2.5 hours a week by our estimate, a month of careful manual checks is roughly 4 to 11 hours of an owner’s attention. On our arithmetic, if your hour is worth more than about $9, Watch costs less than the hours it replaces, before counting the pages it scores and its proposed edits.

Plus lists scheduled tasks and projects among its features. OpenAI’s pricing page does not present them as a brand-tracking tool, and we have not tested them for this job, so we make no claim either way.

How do you check ChatGPT by hand, properly?

A useful manual record begins with 10 written questions and an explicit location, then preserves every answer and its settings. Repeating that record helps distinguish patterns from isolated responses. OperStack automates its own weekly question set; the API method and the ChatGPT website can still return different answers.

1. Write ten buyer questions. Use the words a customer types, not your marketing language. “Which dentist near Brighton takes emergency appointments on Saturdays?” tells you more than “top dental practice”. For a local business, put the city, neighbourhood or postal code into the question; OpenAI’s own search help recommends exactly that for local results.
2. Take yourself out of the answer. Start a temporary chat and choose “Unpersonalized” before typing. OpenAI says that mode leaves out your existing memories, custom instructions and plugins, and temporary chats neither create nor update memories. Signing out also works, with the limits shown in the table further down.
3. Make ChatGPT search. ChatGPT may search on its own when a question needs current information, but you can force it with Search in the tools menu or by typing ”/”. Forcing it keeps your runs comparable.
4. Log four things per answer: the date, the question, every business named in order, and whether you were among them. Note the cited sources too; OpenAI warns that citations “can be incomplete, outdated, or incorrect”, so open the ones that matter.
5. Repeat each question three to five times with “Try again” or a fresh temporary chat, then count the share of answers that named you. The SparkToro authors argue that a visibility percentage across many runs is a more reasonable measure than a position in one list.
6. Confirm you are eligible at all. OpenAI says a site must allow OAI-SearchBot to be eligible for ChatGPT search. Open yourdomain.com/robots.txt and look for a rule that blocks it.
7. Repeat next week with identical settings: same questions, same mode, same location. That makes later observations easier to compare.

No source we found measures how long this takes. Our own estimate, not a measurement: 2 to 3 minutes per question for one unpersonalised, searched answer that you read and log, so 20 to 30 minutes for one pass of ten questions. With three to five repeats, roughly 1 to 2.5 hours a week, before comparing rivals or deciding what to fix.

That time is the heart of the comparison. Watch automates its own recurring question record each week: fixed questions, search on, no personal history, names logged, rivals counted. The manual version works. It simply asks you to be a very consistent research assistant every Monday, with no end date.

Why does ChatGPT give a different answer every time I ask?

Because the system is built to vary, and lists of brands vary most. OpenAI’s API documentation says Chat Completions “are non-deterministic by default”, and a fixed seed yields only “(mostly) deterministic outputs”. For the ChatGPT app, OpenAI publishes no figure for how often the named businesses change.

Independent researchers have measured it. SparkToro and Gumshoe, in work by Rand Fishkin and Patrick O’Donnell, published their study on 27 January 2026: 600 volunteers put 12 prompts through ChatGPT, Claude and Google’s AI Overviews and AI Mode 2,961 times, in November and December 2025, on default settings. They found a less than 1 in 100 chance that ChatGPT or Google’s AI returns the same list of brands in any two of 100 responses. The same order came up about once in 1,000 runs.

The practical reading: one chat can show you absent when you usually appear, or present when you rarely do. Only a tally over repeated, identical runs is worth writing down, and keeping that tally going is the job Watch takes off your desk.

Does ChatGPT show me the same answer my customers get?

Not necessarily, and the business owner is the person most likely to see a skewed answer. If you have spent months talking to ChatGPT about your own company, it may know things about you that a buyer’s ChatGPT does not.

OpenAI’s memory FAQ says that with memory on, ChatGPT can draw on past chats, saved memories, custom instructions, files in Library and connected apps such as Gmail, and that “ChatGPT decides which available information is relevant to a response.” Memory reaches search as well: OpenAI says ChatGPT may use relevant saved memories when rewriting a search query, and saved location details may influence results. Custom instructions apply to every chat at once and exist on all plans, up to 1,500 characters on Free and Go and 5,000 on Plus and above.

Location is the part you cannot switch off. OpenAI says ChatGPT may infer an approximate location from your IP address, at country, state or city level, and that a VPN or network location may affect it. Device location sharing is optional and off by default. ChatGPT also rewrites your question into queries for its search partners; OpenAI’s own example turns “restaurants near me” into “top restaurants San Francisco”. So an owner in one suburb and a buyer in the next can see different shortlists for the same words.

How much personalisation changes which businesses get named is not stated by OpenAI. What is clear: an Unpersonalized temporary chat removes your history but still carries your IP location, so ask from where your buyers are, or name the place in the question. Watch removes the history problem entirely, because it never runs inside your account.

Can I automate ChatGPT checks myself?

A custom script should use the API. Built-in ChatGPT tasks are a separate option, and we have not evaluated their suitability for this workflow. OpenAI’s Pro tiers help page says that under its Terms “automatically extracting data” is prohibited, so a script that drives the ChatGPT website is off the table. The permitted route is the OpenAI API, billed separately from any ChatGPT subscription.

The API has its own model and search-tool charges, separate from a ChatGPT subscription. Check the current official price for the chosen model and log actual usage before estimating a monthly bill. A buyer location can be supplied to the search tool, but API answers should still be recorded as a separate series.

Our estimate, not measured: ten questions a week is about 43 calls a month, roughly $0.43 in search-call fees plus tokens. One answer can trigger more than one search call, and OpenAI does not state token counts per call. The fees are trivial. The real bill is developer time: building the script, pulling business names out of free text, storing the history, telling rivals from directories, scoring your pages, writing fixes, then keeping it all working as models change. OpenAI does not state whether API answers match what a signed-in ChatGPT user sees. Watch is that whole pipeline, already built, for $35 a month.

What does each option tell a business owner on a Monday morning?

A manual Monday produces transcripts only if someone performs the work. Watch produces 1 scheduled email with its question results, page score and proposed edits. The owner still reviews the facts before publishing a paragraph, and a movement in mentions needs interpretation alongside the actual questions that changed.

OperStack Watch keeps the recurring record going when the owner has other work to do. Its 10 questions are followed week by week, with the names returned in the answers, 3 rivals measured alongside the business and a fresh sample of page findings. The later reports compare observations, while the proposed wording supplies an editing task. Manual chats can remain useful for exploring a specific customer question or investigating a source. Keep those exploratory conversations outside the fixed series because changing the wording can change the result. The report is also a record for the site editor: a page address, a dated finding and a suggestion that can be checked against the current offer. Publishing remains a separate step. If the business cannot assign that step to someone, receiving more reports will not by itself repair the website or establish that a customer saw the business in an answer.

1. Preserve manual chats as dated exploratory notes.
2. Choose a Watch page finding you can act on.
3. Record approval and compare later fixed-question observations.

When you need a different kind of tool

ChatGPT Free is useful as a general assistant at $0, including discussion, research and drafting. An owner who needs one exploratory conversation may have no reason to subscribe to monitoring yet. Watch becomes useful when repeating the observation and turning page findings into reviewed edits is a recurring business task.

1. Use a conversation to explore unfamiliar customer language before deciding which questions matter to the business.
2. Open cited pages yourself when researching a specific recommendation. An assistant’s explanation of its answer may be incomplete or unsupported.
3. Ask for help with unrelated work, such as preparing an email or organising notes. A monitoring subscription does not replace that general assistant.

For Google exposure, use the records described in our Search Console comparison. For someone to publish the work on your behalf, the agency comparison covers responsibilities.

Clinic, agency, SaaS company or online shop: which approach fits?

The choice depends on who owns the weekly task across 4 common business situations. A clinic may need a short local record; an agency must repeat work across clients; a software team may build automation; a shop needs careful product coverage. Watch offers a fixed, narrower scope for all four.

1. A clinic

Patients ask local questions, so location matters. Try the manual method once from inside your catchment area, with the suburb in the question. If you are missing, run the free check to inspect possible page-level issues, then start Watch to learn week by week whether the paragraph you pasted moved you onto the list. By hand, that a comparable record requires repeated observations with consistent settings.

2. An agency with ten clients

Manual checking breaks at this size. Ten clients at 1 to 2.5 hours a week each, by our estimate, is 10 to 25 hours of staff time weekly before a single client report is written. Watch for ten clients costs $350 a month at list price, by our arithmetic, with a separate weekly email per client that you can forward as it stands. From Claude or Cursor, the free MCP server puts two, three or four client sites on the same 100-point scale at once.

3. A SaaS company

A team with engineers could build the API route for well under a dollar a month in search fees for ten questions, plus tokens. The fees are not the cost; the parser, the history, the rival sorting, the page scoring and the fix writing are. Watch ships all of that. Developers can also add the free MCP server to Cursor, to Claude, or to ChatGPT on a paid plan with developer mode, and score a page before it goes live.

4. An online shop

Product questions and available stores vary by market, so keep the intended location with the question. A few manual chats from one location will not reveal that spread. Watch keeps ten fixed questions under weekly observation, and the $149 audit reads 100 pages and five rivals in one pass, which suits a large catalogue.

How do you move from manual checks to OperStack?

Keep your existing manual observations when starting Watch, but mark its first 10-question run as a new series. Different questions, locations and collection methods cannot form one continuous percentage. The transition below preserves useful history while separating manual exploration from the recurring report you receive by email.

1. Keep your spreadsheet if you have one. Its earliest rows are your baseline from before any fix.
2. Run the free check. Compare its three answers with your own chats; any differences are a live demonstration of the variance described above.
3. Start Watch; the first 7 days are free and a card is taken at checkout. Your first Watch email shows the ten questions, the names ChatGPT and Perplexity gave for each one and the fixes to make.
4. Paste the first fix, then wait for the next weekly email rather than re-asking the same afternoon. Given how much single answers vary, a same-day recheck proves little.
5. Keep manual chats for the questions Watch reports as lost: ask why the winner was chosen and open its sources.

For a free tool that reports on your brand rather than your pages, see our HubSpot AI Search Grader comparison; for a paid tracker, see Otterly.

What our early sample can and cannot tell a manual checker

Our early sample of September 2026 attempted 511 domains and produced scores for 496. The remaining 15 did not produce a score in that run. That distinction matters when an owner checks manually too: an unavailable result must remain unavailable rather than being recorded as a zero or a failed recommendation. Full study: 731 checks.

Source: OperStack free check, early sample of 25 September 2026, covering small businesses in the United States, United Kingdom, Australia and Canada. The study concerns website readability. It is not a survey of customers, a ChatGPT recommendation experiment or an estimate of the probability that a business wins a lead.

Apply the same distinction to a spreadsheet of manual chats. If the service did not answer, record an error. If it answered without naming a business, record that outcome separately. If the response was about the wrong city, preserve the response but flag its scope instead of silently changing the question and pooling the result.

A single example cannot establish how frequently an error occurs. The benefit of a written procedure is that another person can see what was counted. Keep the original response with the row and add any judgement about competitor identity in a separate column.

Our own-site examples provide context for how we use the tools. Your business needs its own observations; neither our scored-site sample nor a published variability study supplies a conversion estimate for your offer.