Your first 30 days of AI search optimization: a practical plan

In your first 30 days of AI search optimization, do four things in order: record a baseline of how AI engines answer the questions your buyers ask, inspect one gap where a competitor is cited and you are not, make one documented change to one page, and review comparable results with the same prompts, engines and window. The goal of month one is not a big visibility jump. It is a trustworthy starting point, one change you can trace, and an honest read of what you can and cannot conclude yet.

In short

• Week 1 is for measuring, not changing: collect answers, then record a baseline with its sample sizes.
• Week 2 is for choosing one gap and one hypothesis, not a list of fixes.
• Week 3 is for making one change, recording the date, and leaving the rest alone.
• Week 4 is for a like-for-like review; "not enough data yet" is a valid result.
• A change made in the middle of month one cannot get a firm verdict from DiscoveredBy by day 30. Plan for a second review.

What is AI search optimization, and where does it start?

AI search optimization is the practice of improving how often, and how accurately, AI answer engines name and cite your brand. Two terms cover most of it. Generative Engine Optimization (GEO) is the work of earning a place among the sources an AI engine draws on when it composes an answer. Answer Engine Optimization (AEO) is the related work of becoming the direct answer to a question rather than one link among several. The glossary entries for GEO and AEO give the short versions.

The month-one work is the same either way. It rests on three definitions from the docs' key terms:

• A prompt is a question you choose to monitor because it is the kind of thing your buyers ask. It is not a keyword.
• A mention is your brand named in an AI answer.
• A citation is your domain linked as a source for an answer.

You can be mentioned without being cited, and cited without being mentioned, so keep them apart in your notes.

Enlarge image

Week 1: how do I set a baseline?

Spend days 1 to 7 setting up the questions and letting data arrive; make no content changes. A baseline records where you stand before you change anything; without it you cannot tell later whether anything moved.

Start with the prompts, written as your buyers would ask them (comparison, "best for" and problem questions), not as a keyword list. For a method, see how to choose AI tracking prompts. The onboarding steps take a website, a business description with countries, your prompts and, optionally, competitors. Add competitors if you can; a visibility number means more next to theirs.

Then wait for the first scan. A scan runs every tracked prompt against every engine available on your plan. It happens at the next daily run after your prompts exist, which is at most a day away, and you do not start it yourself (your first scan). Citation numbers appear first and mention-based numbers a little later; until then a number reads as no data, not zero.

When answers are in, open Overview and read it in the order the docs suggest (reading your first results):

1. Read the headline and Visibility first. Visibility is the share of analysed answers that name your brand; it answers the most basic question, which is whether engines name you at all.
2. Then read Share of voice, Avg. position and Cited, each of which asks something narrower.
3. Check the notices under the headline. "Failed runs in this window" and "Brand figures use N of M answers" tell you how complete the baseline is.
4. Look at sample sizes. A figure built from fewer than 30 observations still shows, but with a "Low sample" label, and you should treat it as provisional.

Enlarge image

A 7, 28 or 90 day window ends yesterday, and each change is measured against the same number of days before it. A brand-new project has no earlier period, so at first you have levels but no changes. Record the levels.

Week 2: how do I choose which gap to work on?

Spend days 8 to 14 picking one gap, one page and one hypothesis. With five changes in a month you will not know which one mattered.

Start from evidence. On Overview, the recent-answers feed shows each answer as "Mentioned #N", "Not mentioned" or "Not analysed yet". Open one where you are not mentioned. The answer drawer marks each source Your page, Mentions you, No mention or Not checked, which tells you whether a cited page names you.

Then use Citation gaps. A prompt becomes a gap candidate only when two things are both true over the trailing window: you appear in under 30% of that prompt's own runs, and at least one citation points at a domain that is not yours. DiscoveredBy compares the cited competing pages with your best matching page and records the specific differences it finds as evidence, such as a comparison table, an FAQ, a statistic or a topic your page does not cover. When no page of yours matches, the recommended fix is to create a page rather than edit one.

For month one, pick a gap that has a page of yours behind it. That is the case Optimizations handles: it reads the pages engines cite instead of yours and drafts a change for the one page of yours. A gap with no page of yours never becomes an optimization; the route there is new content instead.

Write your hypothesis as a sentence you could be wrong about: "Competitor pages for this prompt include a comparison table; ours does not; adding one may make our page a more useful source for this question." It is a hypothesis to test. Nothing in the docs says a table, an FAQ or any other format will earn a citation.

For a longer treatment of choosing among gaps, see find the citation gaps that deserve your next content update.

Week 3: how do I make one documented change?

Spend days 15 to 21 making the one edit, recording it, and then leaving that page and its prompts alone. The point of a documented change is that on day 30 you can name what changed, when, and for which prompts.

Open the Optimizations screen. It opens on the highest-scoring page still waiting on a fix, with a "Write the fix" button if nothing has been drafted. Some conditions apply before a draft can exist:

• The feature depends on your plan; see plans and limits.
• Your target page and the top cited source pages must have been read (scraped) first; if not, they are queued and you try again later.
• A proposed change is kept only if one of its evidence quotes is a real phrase, at least sixteen characters, found in the source page's text. That checks the evidence, not the draft body, so read the draft as a starting point, not text verified word for word.

Read the draft, edit it, and publish it on your own page yourself; nothing reaches your site without you. Then choose Mark applied, which records the trailing month of visibility for the fix's prompts and shows the outcome as "pending more data". Also write the date, URL and a one-line description in your own log.

Enlarge image

Keep the change small enough to describe in a sentence, and do not edit the review-set prompts in the same period; a changed prompt set changes the population you are comparing.

Week 4: how do I review results fairly?

Spend days 22 to 30 comparing like with like and writing the verdict, including the possibility that there is no verdict yet. A fair review uses the same prompts, the same engines and a window that starts after the edit went live.

Use the fix's own prompts, not the project-wide headline. On Optimizations, choose Measure to compare current visibility for those prompts with the baseline. Then apply these rules for how to read the label:

• Under a week since applying: checking always reads "pending more data".
• From a week to a month: the outcome stays "pending" unless visibility for those prompts has already jumped by ten points or more, in which case it reads "positive" early.
• At the month mark: "positive" needs five points or more, "negative" needs five points or more the other way, and anything smaller reads "neutral", which means nothing definitive happened, not necessarily that the fix failed.
• No completed answers for the fix's prompts in the period means no score is recorded and the outcome stays "pending more data".

The check's baseline is re-measured each time from the fix's prompts in the 30 days before you marked it applied. If an engine started or stopped answering part way through, only engines measured in both periods are compared. The metrics reference covers the "no data is not zero" rule and the 30-observation floor.

Put this together and the timing becomes clear. If you apply the change on day 16, a day-30 check is two weeks in. A firm month-mark verdict arrives around day 46. So your day-30 review is a first read, and your log should say so. The Citation gaps queue has its own checkpoints at 7, 14 and 30 days after you mark a fix done, with different rules; see Citation gaps.

For a fuller before-and-after method, see did your content update help?.

The 30-day checklist

Copy this into your team's tracker. Every item is something you do, not something the product does for you.

DAYS 1-7: BASELINE (no content changes)
- [ ] Write your starter prompts in buyers' words (comparison, "best for", problem)
- [ ] Add the project's countries and, if possible, your main competitors
- [ ] Wait for the first daily scan; do not judge mention numbers on day 1
- [ ] Read Overview: headline, Visibility first, then Share of voice,
      Avg. position, Cited
- [ ] Check "Failed runs" and "Brand figures use N of M answers" notices
- [ ] Note every "Low sample" label (fewer than 30 observations)
- [ ] Fill in the baseline log below and save the date

DAYS 8-14: INSPECT ONE GAP
- [ ] Open 3 to 5 answers that do not mention you; note the sources
- [ ] Open Citation gaps; keep only gaps with a page of yours behind them
- [ ] Choose ONE gap, ONE page, ONE hypothesis
- [ ] Write the hypothesis so it could be wrong
- [ ] Note the prompts the gap affects (this is your review set)

DAYS 15-21: MAKE ONE DOCUMENTED CHANGE
- [ ] Confirm the target and cited source pages have been read; retry if not
- [ ] Generate the draft on Optimizations ("Write the fix")
- [ ] Review the draft and its evidence quotes; edit for accuracy and voice
- [ ] Publish the change on your own page
- [ ] Mark applied; write the date, URL and one-line edit in the change log
- [ ] Freeze: no other edits to this page, no edits to the review-set prompts

DAYS 22-30: REVIEW LIKE FOR LIKE
- [ ] Press Measure on Optimizations; use the same prompts, engines and a window after the edit date
- [ ] Label anything under 30 observations provisional
- [ ] Record the outcome: improved, no change, or insufficient evidence
- [ ] Record what else changed in the period (content, PR, launches)
- [ ] Decide: keep monitoring to the month mark, revert, or pick gap 2
- [ ] Schedule the firm review (30 days after the edit date)

And a log to keep alongside it:

A worked example

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to mid-sized legal and compliance teams. It tracks 12 prompts and one competitor, Brieflane.

Day 7 baseline. Overview shows 40 analysed answers. Quillstone is named in 10 of them, so Visibility is 10 of 40, or 25.0%. Brieflane is named in 22 of 40, or 55.0%. Of 48 collected answers, 6 cite a Quillstone page, so the Cited rate is 12.5%. Both denominators (40 analysed answers and 48 collected answers) are at or above the 30-observation floor, so nothing is labelled low sample.

Day 12 gap. Reading answers where Quillstone is not mentioned, the team sees the same page cited for "best document review software for compliance teams". Citation gaps flags that prompt cluster: Quillstone appears in 2 of 12 runs for that prompt cluster (16.7%), under the 30% threshold, and a Brieflane comparison page is cited. Quillstone does have a matching page. The hypothesis: the cited page has a side-by-side comparison table and Quillstone's page does not.

Day 16 change. The team generates a draft on Optimizations, checks the evidence quote against the source page, rewrites the table in Quillstone's own words and publishes it. They mark it applied and freeze the page.

Day 30 review. The team chooses Measure. For the same cluster, Quillstone is named in 3 of 14 answers since the edit, or 21.4%, against 16.7% in the 2 of 12 the team recorded on day 12. That is a rise of 4.7 points on 14 answers. (The product re-measures its own baseline from the 30 days before Mark applied, so its figure may differ from the team's day-12 note; the team keeps both.) The docs' rule says a change of under ten points inside the month reads "pending". The sample is also below 30 observations, so the figure is provisional. Their verdict: insufficient evidence. They record it as such, keep the page frozen, and schedule the firm review for around day 46. They do not claim the table worked, and they do not claim it failed.

Common mistakes and what this plan cannot tell you

• Changing many things at once. Any movement becomes unattributable.
• Treating a first look as a baseline. Early data can be thin, partly analysed or missing an engine. Note sample sizes.
• Reading the project-wide headline as the verdict. It moves for reasons unrelated to your one edit. Review the fix's own prompts.
• Assuming a format will earn a citation. A drafted table, FAQ or fresher date is something to test.
• Assuming an observed change proves a cause. Even a ten-point difference sits alongside every other change in your program and in the engines. It is an observation, not proof.

Frequently asked questions

How many prompts should I track in month one?

The docs do not give a required number. Track enough that each review set has a meaningful number of answers, remembering that answers for one prompt accumulate daily across engines, and that figures under 30 observations are provisional. Your plan sets a limit on prompt slots; see plans and limits.

Should I make several changes to move faster?

Not in the first cycle. A single change can be attributed and reviewed; several cannot. Once you have one clean loop, run the next gap the same way.

Can I judge the change after two weeks?

Only as a first read. Under a week since marking a fix applied, the check always reads "pending more data". Between a week and a month it stays pending unless visibility jumped ten points or more. A firm reading needs the month mark, and even then a "neutral" means nothing definitive happened, not that the change failed.

What if there is no gap with a page of mine behind it?

Then the fix is a page that does not exist yet. Accepting a citation gap creates a candidate topic on the Articles screen, and nothing is written until you pick it there. That becomes your month-one change.

Do I need to know GEO and AEO before I start?

No. The plan works from what your own answers show. GEO versus SEO covers what carries over from search engine optimization.

What if visibility goes down after my change?

Check sample size, failed runs and engine coverage first. A drop can come from a smaller sample, a different engine mix or ordinary variation between days. Treat it as an observation to investigate, not evidence that the edit hurt.

Next step: run your first loop in DiscoveredBy

The steps above map onto Overview, Citation gaps and Optimizations. Optimize describes how gaps become page-level fixes, and the docs on reading your first results and optimizations cover each screen. To start your baseline, sign up or sign in and add your first prompts.