Omni Eclipse
Omni Eclipse
Guides

Set a Baseline Before You Start AEO: What to Record on Day One

Ashur Homa
Ashur Homa
·October 20, 2026·12 min read·
Set a Baseline Before You Start AEO: What to Record on Day One

Auditing your AI search visibility starts with a fixed list of 10 to 15 real buyer questions, tested across ChatGPT, Google AI Overviews, Perplexity, Gemini and Claude, with every answer logged the same way before you touch a single page. That one session, done on day one, is the baseline every later result gets compared against.

Skip it and you will not know, in three months, whether anything you did actually worked. You will have a feeling that things seem better, which is not a measurement.

✅Day one checklist

Before you write, publish or fix anything, record this once:

  1. 10 to 15 fixed prompts, written the way a buyer would actually type them, not keyword phrases.
  2. 5 engines: ChatGPT, Google AI Overviews, Perplexity, Gemini and Claude.
  3. Logged out, every time. A signed-in account has seen you discuss your own business and will mention it. That is memory, not visibility.
  4. Two columns per answer: was the business named, and did its domain show up as a cited source.
  5. One screenshot or saved transcript per answer, dated, kept somewhere you can find again in eight weeks.
  6. A fixed re-test date, two to four weeks out. Not "soon."

Why does a baseline matter before you start AEO?

A baseline matters because without one, you cannot tell a real change from the normal day-to-day noise in how these engines answer. AEO work takes weeks to show up, and if you start measuring only after you have made changes, you have no "before" to compare the "after" against, so every result is a guess dressed up as a fact.

This is the single most skipped step in AEO, and it is skipped for an understandable reason: a baseline produces no content, fixes nothing, and feels like delay. But it is the only thing standing between "we think it's working" and "here is what changed, and here is what it was before." An agency, or a founder doing this themselves, that starts writing content on day one with no recorded starting point is the agency you cannot later ask to prove anything, because there is nothing to prove it against.

800M+
ChatGPT weekly users
OpenAI, via TechCrunch, Oct 2025
2B+
Monthly AI Overviews users
Google, via TechCrunch, Jul 2025
25%
Marketers tracking nothing at all
Page One Power, Mar 2026
<1%
Exact answer repeat rate, run to run
SparkToro & Gumshoe.ai, Jan 2026

The scale is why this matters at all. ChatGPT has more than 800 million weekly users (OpenAI, via TechCrunch, October 2025), and Google's AI Overviews now serve more than 2 billion monthly users across over 200 countries (Sundar Pichai, Alphabet's Q2 2025 earnings call, reported by TechCrunch, July 2025). Your buyers are already asking these systems questions about your category whether you are watching or not. A Page One Power survey of 600 marketing professionals found that 25% are not tracking AI visibility at all, and another 12% did not know tracking it was possible (Page One Power, March 2026). Most businesses starting AEO are starting from zero records, not just zero optimization.

How many prompts do you need for day one?

Ten to fifteen is enough for a first baseline, and more is usually worse, not better. A larger set produces a tidy average that nobody has time to read, and reading the actual answers is where the useful information lives.

Choose the prompts from real buying language, not from keyword research. Pull your last 30 to 50 inbound inquiries or sales calls and find the handful of distinct questions underneath them. A plumbing business's customers do not search "plumbing services," they ask "who can fix a burst pipe on a weekend near me." A B2B software buyer does not ask about your category name, they ask which tool handles a specific integration or team size.

For day one, aim for this mix:

Four on the decision itself. How to choose a provider in your category, what it typically costs, what to look for, and the red flags to avoid.

Four to six on the specific job your best customers actually arrive with. Not your category description: the real situation that brought them to you.

Two to three on your location or segment, phrased the way someone would say it out loud, not the way a directory listing would.

One or two on your brand name, to see what the models currently say about you when they do recognize you. They will happily repeat a price you changed two years ago or a service line you no longer offer.

Write the list down and do not touch it. Changing prompts between the baseline and the re-test changes the result without changing anything real about your visibility. If you want to track a new question, add it as an addition, not a swap, so the original set stays comparable over time. Our fuller breakdown of how many prompts to track goes deeper into choosing and sizing the set once day one is behind you.

Which engines do you need to test on day one?

Test all five of ChatGPT, Google AI Overviews, Perplexity, Gemini and Claude, because each one draws on different sources and will show you a different business. Testing only one tells you about that one engine, not about your AI visibility.

ChatGPT is the obvious starting point given its reach, but Google AI Overviews reaches a comparable audience from inside ordinary search, so a business invisible there is invisible in a channel most of its customers never think of as "AI search" at all. Perplexity is worth the extra few minutes because it shows a visible numbered source for nearly every claim, which makes it the most diagnostic engine to read: you can see exactly where an answer about you, or about a competitor, actually came from. Gemini and Claude round out the set, and each occasionally surfaces a business the other three miss entirely, usually because of what each model happened to be trained or retrieved on.

If you genuinely cannot do all five on day one, do ChatGPT and Google AI Overviews first and add the rest within the week. Do not permanently drop any of them: a gap in one engine is information, not noise, and you only learn that by looking.

What exactly counts as a mention, and what counts as a citation?

A mention is your business named in the body of the answer. A citation is your domain appearing as a source the answer links back to. Record these as two separate columns, because they measure two different things and point at two different fixes.

Being cited means an engine used your content as reading material to build its answer. Being named means it decided your business was worth recommending to the person asking. These overlap less than most people expect. A business can be cited constantly and named rarely, which means its content is doing its job but its presence elsewhere, reviews, directories, other people's pages, is not. A business can be named occasionally with no citation at all, which usually means a model recognizes the brand from general knowledge but is not pulling anything current about it.

Collapsing both into a single "visibility" number hides which situation you are actually in, and the two situations need close to opposite work. This is also the first thing to check in any report an agency hands you: if it gives one blended figure, ask how it defines a mention, because the definition decides the whole number.

For each of your prompts, on each engine, record:

  • Was the business named in the body of the answer: yes or no.
  • Did the business's own domain appear as a cited source: yes or no.
  • What specific claim was made about the business, if any (price, location, service, differentiator).
  • Which competitors appeared, and in what position.
  • Whether anything stated was factually wrong.

How do you log the results so they are actually usable later?

Log every answer the same way, in one place, dated, with the full text saved, not just a yes or no. A spreadsheet with one row per prompt per engine per date works fine; the format matters far less than the consistency.

Minimum columns: date, engine, prompt (exact wording), named (yes/no), cited (yes/no), competitors mentioned, specific claims made about you, and a link to a saved screenshot or copied transcript. Save the full answer text somewhere retrievable, not just your interpretation of it. In six weeks you will want to reread the actual wording, because the thing that changed is often subtler than "we went from invisible to visible": it is a claim getting corrected, or a competitor dropping out of one specific answer.

Always test logged out, in a fresh or private session. A signed-in account has prior conversation history, and these systems increasingly use that memory. If you have ever discussed your own company with your logged-in ChatGPT account, it may mention you later not because you are visible to the public, but because it remembers you asking about yourself. That single mistake is the most common reason a business believes its AI visibility is better than it is.

Run each prompt more than once, on different days rather than back to back. These systems are not fully deterministic even with identical wording: a study by SparkToro and Gumshoe.ai, which ran 12 brand-recommendation prompts roughly 2,961 times across major AI platforms, found that the exact list of names an engine returns for the same prompt repeats less than 1% of the time run to run (SparkToro & Gumshoe.ai, January 2026). The specific wording of an answer will vary. What stays comparatively stable, across enough repeats, is whether a brand shows up in the pool of names at all, which is exactly why a single run on day one should be treated as a rough first read, not gospel, and why a second pass two or three days later strengthens it considerably. For a practical walkthrough of the manual testing mechanics on each platform, see our guide to checking your business's AI search visibility.

What mistakes turn a baseline into a wasted afternoon?

The most common mistake is testing logged in, which quietly inflates the result with memory instead of real visibility. The second most common is treating one run of each prompt as final, when a single run on a non-deterministic system is a sample, not a measurement.

A few others worth guarding against on day one:

Changing the prompt wording "to get a better answer." The point of a baseline is comparability over time, not getting the most flattering result today. If a prompt returns nothing useful, that is data. Rewrite it for the next period, not mid-session.

Testing only your brand name. A prompt that includes your company name tests recognition, not discovery. Someone who has never heard of you will never type your name. Most of your set should be the questions a stranger asks.

Skipping the losing questions. It is tempting to test only the prompts you expect to do well on. The informative prompts are the ones where you suspect a competitor already wins, because that answer tells you something you did not already know.

Recording a score instead of the answer. "60% visible" means nothing without the underlying text. Keep the actual wording of every answer, because the diagnosis lives in the specific claim, not the percentage.

Not fixing a re-test date. A baseline with no planned second measurement is a one-time curiosity, not a baseline. Put the date in a calendar before you close the spreadsheet.

How often should you re-run the baseline after day one?

Re-run the full set every two to four weeks, not daily and not quarterly. Weekly testing mostly shows you the normal variance described above and will talk you out of an approach that is actually working. Quarterly testing is too slow to catch a real decline before it compounds, and too slow to know which specific change moved a result.

Two to four weeks gives enough separation for genuine change to show through the noise, while staying frequent enough that you can connect a shift to something you actually did. Keep the same prompts, the same engines, and the same logged-out method every time; the method is what makes the second measurement comparable to the first rather than a new, unrelated data point. Once you have two or three cycles of data, you will also have a sense of your own baseline variance, which tells you how big a change needs to be before it is worth reacting to. If you want a fuller framework for what to track once the baseline is running, including citation accuracy and conversion, see how to measure AEO results.

What does a good day-one result actually look like?

A good day-one result is not a high number. It is a complete, honest record of where you currently stand, read in full, with the pattern visible across prompts and engines. Most businesses starting this process are named on a handful of their ten to fifteen prompts and absent on the rest, which is normal, not a failure.

What you are looking for on day one is not a verdict but a shape. Named on narrow, specific questions but absent on broad category ones is the healthiest pattern to start from: broad questions have more competitors and older, deeper sources behind them, so narrow wins tend to lead to broad ones over time. Named on broad questions but absent on narrow ones is less common and usually means the models have heard of your brand in general but do not know specifically what you do, which is a positioning gap showing up in the measurement. Visible on one engine and invisible on the rest is extremely common and mostly reflects which sources that particular engine leans on; check what got cited where you did appear, then look for an equivalent source feeding the engines where you did not.

None of this is diagnosable from a single visibility percentage. It is only visible if you read the actual answers, which is the entire argument for keeping the prompt set small enough to read on day one and every re-test after it.

Get your real baseline, not a guess

A free AI visibility audit on your own buying questions, across all five answer engines, logged the right way from day one.

Book your audit
Ashur Homa
Written by

Ashur Homa

Growth @ Omni Eclipse

Built and scaled a digital brand to $100M+ in sales with zero ad spend. Has helped businesses generate millions through AI go-to-market strategy. Leads growth at Omni Eclipse.

Connect on LinkedIn

What we have measured so far

82,000+
AI answers measured
Across every tracked client
1,000+
Buyer prompts tracked
The questions their customers actually ask
5
Answer engines
ChatGPT, Google AI Overviews, AI Mode, Perplexity, Gemini
143
Citation wins
Times we moved a brand into an answer

What the engines actually say

Quoted exactly as returned, on a real buying question, naming a real client.

Google AI Overviews answering "What is the ROI of connecting ALM, DevOps and ITSM tools?"
Fast time-to-value: Integration platforms (such as OpsHub Integration Manager) connect existing tools in weeks, avoiding multi-year migration projects. Minimal disruption: Development, QA and IT support teams keep their preferred native interfaces.
How OpsHub got there
Google AI Overviews answering "What is the best ALM migration tool for software teams?"
The best Application Lifecycle Management (ALM) migration tool depends on your specific platforms, but OpsHub Migration Manager is widely considered the top purpose-built solution. It specializes in high-fidelity data migration, preserving complex work item hierarchies, history, and attachments with minimal downtime.
How OpsHub got there

In their words

As a newer ServiceNow partner, we're building our presence in a space where established firms have had years to develop their digital footprints. We knew our SPM expertise was there; the opportunity was to make it easier to find in AI search and chat. Omni Eclipse helped us focus that effort, and we're now seeing CoreX cited much more consistently in the SPM answers we track. It's exciting to see our visibility begin to catch up with the work our team does every day.
Devon Clarke
Head of Global Marketing at CoreX
See the numbers
What we do is genuinely hard to explain in one line, and for a long time AI was explaining it for us using everybody else's pages. Omni Eclipse worked out which questions our buyers were actually asking and got our own answers in front of them. The people who reach us now already understand what we do before the first call, which changes that conversation completely.
Aparna Garg
Director of Product and Marketing at OpsHub
See the numbers
Brokers were asking AI which platform to use and getting an answer assembled out of everybody else's marketing. Omni Eclipse worked out exactly which questions our brokers were asking, then got our own explanation into the answer. AI does not just name us now, it describes what we actually do, and the brokers who reach us already understand it.
Michelle Peters
Marketing Manager at COG Aggregation
See the numbers

Why choose Omni Eclipse

Doing it in-houseA generalist agencyOmni Eclipse
What gets measuredRankings and sessions, which no longer describe how buyers arriveTraffic, with AI search reported as a line itemWhich prompts name you, on which engine, tracked daily across five
What you see between reportsWhatever someone has time to pull togetherA monthly deckA live portal. The same view we work from, open to you
Who does the workA marketer adding this to an existing jobAn account manager briefing a content teamThe people who built the measurement, working the accounts
Where the work happensYour own website, which is the easier halfYour own website, at volumeYour site and the third-party sources engines read before answering
How long before you knowUnclear, because nothing is being tracked from a baselineTwo to three months to a first reportA measured baseline in week one, movement visible weekly

Free AI visibility audit

Is your business visible to AI search?

Find out where you stand across ChatGPT, Google AI Overviews, Perplexity, and more. Then see what it takes to become the recommendation.

Get Your Free Audit