Blog · Running the work

The first 90 days of AEO work

Published 7 May 2026 · Updated 12 October 2026 · 8 min read

The first 90 days of AEO work
Photograph: Zhuo Cheng you / unsplash

The short answer: A first quarter buys you a baseline, the access and identity fixes, and one honest re-measurement. Weeks one and two are measurement only. Weeks three to six ship the fixes that depend on nobody outside your company. Weeks seven to ten wait, because refetch is slower than deployment. The last fortnight re-runs the same prompts and tells you which of the three layers is still broken.

Ninety days is an awkward length. Long enough that an access problem will be fixed and visible, short enough that anything depending on other people's websites will still be in flight. Teams get into trouble by expecting those two to arrive together.

Weeks one and two: measure before you touch anything

The single most expensive mistake in a first quarter is shipping a fix before recording what the engines said beforehand. Once you have redeployed, the comparison is gone. You cannot reconstruct a baseline later, because the answer you are comparing against was generated on a day that no longer exists.

A baseline needs three properties and they are all boring. The prompt list is fixed and written down. Each prompt runs several times on each engine, not once, because in our own audit runs the same prompt minutes apart has come back with a different source list on the same engine. And every run is stored with the citation list attached, so a finding can be argued with rather than believed.

This is what an AEO audit should contain, and it is the first fortnight's only deliverable. Resist the urge to also fix the three obvious things you spotted on day two. Write them down instead. The sample audit report shows the shape the evidence takes.

One practical note on the prompt list. Build it from the questions a buyer asks near a decision, not the category explainers that flatter you. An engine explaining what a category is often names nobody. An engine answering who to use for a specific job tends to name a handful of companies, and that is the surface you are competing on.

Weeks three to six: the fixes nobody else has to approve

Now ship, in this order: access first, then identity, then the pages themselves.

Access means the engines can fetch and parse the page. Check robots.txt rules per agent rather than in aggregate, because the agents do different jobs. OpenAI's documentation separates OAI-SearchBot, which surfaces sites in ChatGPT search, from GPTBot, which collects training data, and from ChatGPT-User, which fetches a page because a person asked for it. Blocking all three is a decision. Blocking them by accident, through a bot rule written before those agents existed, is one of the findings we hit most often in the audits we have run. That is a statement about the sites that come to us for an audit, not a measurement of the web.

Here is a check worth running this week. Pull a fortnight of server logs, split the hits by user agent, and then verify them against the published IP ranges rather than trusting the string. OpenAI's bot documentation links a separate JSON range file per agent, and Perplexity's links one for PerplexityBot and one for Perplexity-User, both checked on 12 October 2026. A user-agent header is a claim anyone can make, so a log count that is not IP verified counts anything that copied the string, and will read higher than verified crawler traffic. The reverse case matters too: a WAF rule that challenges an unverified agent will also challenge the verified one if the rule keys on something else, such as a datacentre ASN.

Identity work runs next because it is slower to settle. Organisation markup, consistent naming across your own properties, the profile links that point back at each other. How much of this actually matters varies by sector, which is the whole argument in AEO by industry: a regulated lender and a kitchenware retailer fail at different layers, and a generic first-quarter plan wastes money on both.

Weeks seven to ten: the wait that looks like nothing happening

This is where first quarters get cancelled. The fixes are live, the dashboard has not moved, and somebody asks what they are paying for.

Three published facts explain part of the delay, and each is specific to the engine that published them. Google's documentation on AI features, last updated 10 December 2025, says there are "no additional requirements to appear in AI Overviews or AI Mode" beyond the usual indexing and snippet eligibility, and that "crawling can take anywhere from several days to several months, depending on how often our systems determine a page needs to be refreshed". That is Google's clock, not a figure for the other engines. Separately, Perplexity's crawler documentation says a robots.txt change may take up to 24 hours to be reflected in its systems. OpenAI's crawler documentation, read on 12 October 2026, gives the same order of magnitude for its own systems: for search results it can take around 24 hours from a robots.txt update for them to adjust. What is common to all of them is that measuring the afternoon you deploy measures nothing.

There is one lever that shortens part of the wait, and it only reaches part of the stack. IndexNow lets you push changed URLs to participating engines, up to 10,000 per POST by its documentation, and its FAQ describes 202 Accepted as submission accepted but not yet processed, which is not the same as indexed. The endpoints IndexNow lists are Amazon, Bing, Naver, Seznam, Yandex and Yep. Neither ChatGPT nor Perplexity is on that list, so there is nothing to push to them. We have not measured how much a push shortens the wait on any engine.

Use these weeks for the work that has a long fuse anyway: the corroboration you do not own. Getting a wrong fact corrected on a third-party profile, publishing the data somebody else would need to cite you, answering the questions in the places your buyers already read. None of it lands inside the quarter. All of it has to start inside the quarter.

Weeks eleven to thirteen: re-run the same thing, change nothing else

Re-measurement is only evidence if it is identical to the baseline. Same prompts, same engines, same number of runs per prompt, same treatment of retrieval on and off. Add a prompt and you have added a confound. A prompt set quietly expanded between runs will tend to report improvement, because the new prompts were picked after somebody knew what was working. We have not counted how often this happens, so treat it as a reason to freeze the set rather than as a finding.

Read the result by layer rather than as one score. Access fixes that worked show up as pages being fetched and cited at all. Identity fixes show up as the engine describing you correctly instead of hedging. Corroboration will mostly not have moved, and that is the expected result at day 90, not a failure. Report it that way before anyone asks.

Expect some noise in both directions. Citations are not durable the way a ranking was, because the answer is regenerated on every ask: in our own re-runs the same prompt has cited a page one week and not the next with nothing changed on the site. We have no public dataset on how long a citation survives, so we will not put a half-life on it. A page winning on one engine and quiet on another is ordinary, not a bug in your work.

What 90 days does not buy

It does not buy a stable answer. The answer is regenerated on every ask, so a single good run is an anecdote and a single bad one is too.

It does not buy attribution. In the client analytics we have been shown, a large share of the traffic that plausibly came from these surfaces arrives with no usable referrer, and we cannot measure what we are not sent. The honest end-of-quarter report connects fixes to mentions, and stops there.

It does not buy a finished programme. What a quarter should leave behind is a prompt set somebody owns, a fixed way of running it, and a written list of what is still broken and why. What an AEO budget actually buys is worth reading before the second quarter is planned, because the second one costs differently from the first. If the terminology is still slippery inside the company, AEO vs GEO vs SEO settles it in one read.

Honest limits

The fortnight boundaries in this post are a working plan, not a measured optimum. We have not run the same programme with a different schedule on comparable sites, so we cannot show that three weeks of baseline or six weeks of waiting would do worse. The three clocks we quote are each one engine's published statement about itself, and we have no published propagation figure for ChatGPT, Gemini, Claude, Copilot or Grok. Where this post says what we have seen, it describes the sites that came to us for an audit, which is a self-selected sample and not a sample of the web. We have not measured how much an IndexNow push shortens the wait on any engine, and we cannot see traffic that arrives without a referrer, so an absence in our logs is our blind spot before it is anybody's evidence.

FAQ

Can we skip the baseline and start fixing? You can, and you lose the ability to attribute any change to any fix. A baseline costs two weeks once. Not having one costs you every argument for the rest of the year.

How often should we re-measure in the first quarter? Twice. At the start and near the end. Weekly re-runs inside a crawl cycle mostly measure run-to-run variance, which then gets reported as progress or decline.

What if nothing has moved at day 90? Check in order: was the page refetched, was it parsed with the answer intact, and does anything the engine trusts repeat your claim. In the stalled quarters we have looked at it is usually the first of those, and server logs will tell you within an hour.

Is three months enough to see revenue? Usually not, and anyone offering that timeline is guessing. Mentions in answers are the measurable outcome at this horizon.

Sources

Related reading

What an AEO audit should containWhat an AEO budget actually buysAEO by industry: where the work differsAEO vs GEO vs SEO: what the three terms actually meanGovernance, and one source of truth for brand factsWhen the model makes something up about you