What to Do in Your First 90 Days of AI Representation Work

A team gets its first real picture of how AI describes the brand. The founding date is wrong on the reference layer. Two engines put a competitor first on the main category question. A third cannot read half the product pages.

Then comes the harder question. What happens on Monday?

AI representation refers to how AI systems describe, position, and recommend a brand when answering buyers’ questions. GEO, or Generative Engine Optimisation, is the discipline of improving it. The first quarter of GEO work is not measured by how much a team ships.

It is measured by whether the team can, at day 90, say what was true at day 1, what changed, and why.

Most first quarters fail one of those three tests.

Why can’t you optimise everything at once?

The first is noise. SparkToro found that fewer than 1% of AI recommendation lists return identical results when the same prompt is run twice. AI answers are probabilistic, so a single check before and a single check after cannot distinguish your work from ordinary variance.

The second is attribution. A team that corrects its entity data, commissions 12 articles, and reopens crawler access in the same three weeks will learn that something moved. It will not learn what.

That matters more in an enterprise than it sounds. At the next budget cycle, the question is not whether AI visibility improved. It is which line item to fund again.

Capacity is the third constraint, and the quietest. A serious tracking programme surfaces 20 to 30 possible Activations a month. No marketing team ships 30 things a month. The work of the first quarter is choosing, not doing.

Days 1 to 30: build a baseline that can detect change

• Measure per engine, never in aggregate. Overlap between the two largest assistants is around 11%. Wikipedia accounts for 47.9% of ChatGPT’s top-10 citations and under 1% of Perplexity’s. A single blended score hides the only differences worth acting on.

• Sample repeatedly, across many phrasings of the same intent. Engines differ enormously in how often they name anyone at all. One analysis of 680 million citations found ChatGPT mentions brands in 0.59% of responses, Perplexity in 13.05%, and Grok in 27.0%. One question is a data point. A measurement needs many.

• Baseline the technical layer first, because it invalidates everything else. One 2026 analysis found technical barriers blocking AI crawlers on 73% of sites. If AI cannot read the content, no other measurement means what it appears to mean.

• Audit the entity data. Category, founding date, ownership, product lines, market presence. This is the factual baseline engines reason from, and errors here propagate into every answer downstream.

• Record competitive standing on real buyer intents, not brand-name queries. Asking an engine about your own brand tells you what it holds. Asking the question a buyer actually asks tells you whether you appear at all.

Sector context belongs in the baseline too. Conductor’s benchmark across 21.9 million searches puts AI Overview trigger rates at 48.75% in healthcare, 25.79% in financial services, and 4.48% in real estate.

A real estate brand will see far less signal per query than a healthcare brand, and should size its sampling accordingly rather than concluding it has no exposure.

One caveat we now build into every engagement. If the intent set expands during the quarter, and it usually does, the day 90 score is measured across a broader, harder set of questions than the day 1 score. Across our client engagements, the headline numbers almost always dip in the first quarter for exactly this reason, and almost none of it is real decline. Treat the first reading as provisional, keep a stable core of intents that never changes, and read the trend against that core.

By day 30, the team should be able to answer four questions per engine. How is the brand described? Which sources produce that description? Where does it stand on the intents that matter? What can AI read, and what can it not?

Days 31 to 60: change the first few things, properly

• Entity and technical changes register fastest, often within weeks, because they change what an engine retrieves rather than what exists to be retrieved.

• Review and directory data follow, since the platforms are indexed frequently and the change is a data correction rather than a body of work.

• Editorial and earned media is a months-long programme. Consistency over time is what gets retrieved, so a single placement will not show up in a 90-day window.

• Community presence is the slowest and the least controllable, because it has to be earned inside conversations the brand does not own.

For most enterprise teams, that ordering points at entity data or technical access. Both are unglamorous. Both are also the cheapest work on the list and the only work that can be finished inside a quarter.

One pattern shows up consistently in our own client data. The brands that complete the most Activations in a first quarter are not the brands with the strongest trajectory. Sequencing against a measured gap beats volume of completed work, every time.

Hold the other five steady, and write down what changed anyway. A product launch, a funding announcement, or a press cycle will move the numbers, and at day 90 you will need to know it happened.

One caution worth building into the plan. Shipping a change is not the same as making one. A corrected reference entry that no engine has re-crawled has not changed anything yet. Recency weighting also differs by engine, so the same edit lands at different times in different places.

Days 61 to 90: verify, then set the cadence

Re-measure with the same protocol. Same intents, same phrasings, same engines, same sampling depth. A baseline is only worth what its repeatability is worth.

Expect movement in the layer you touched and drift everywhere else. Search Engine Land reported that between 40% and 60% of the domains AI systems cite turn over from one month to the next. A portion of any change you see is the market moving, not the brand.

This is why the plan needs a control. Keep a set of intents you deliberately did not work on, and read your results against their movement rather than against zero.

Then set the operating rhythm the rest of the programme runs on: measurement monthly, prioritisation monthly, a board-level view quarterly.

The cadence is the actual deliverable of a first quarter. The single change is the proof that the cadence works.

What does good look like at day 90?

Four things, and they are more modest than most kick-off decks promise.

• A baseline per engine that can be reproduced by someone else on the team.

• One completed change with evidence that it registered, read against a control.

• A ranked queue for the next two quarters, ordered by expected impact rather than by ease.

• A reporting cadence the board has seen once and understood.

What is not realistic at day 90: leadership on every intent, a settled source mix, or a single number that will still mean the same thing in a year.

This is not a hedge. Across the enterprise engagements we run, the first quarter is where the measurement gets built. The real movement in share of voice shows up in months 4 to 6. Teams that expect the curve a quarter earlier tend to conclude the programme has failed at precisely the point it starts working.

An enterprise team presenting at the end of quarter one is presenting a method and a first result. That is a stronger position than a large improvement with no attribution behind it, and it survives the follow-up questions.

Four ways a first quarter gets spent without being used

The first 90 days aren’t about chasing every signal or building along list of initiatives. They’re about establishing a reliableapproach: a repeatable baseline, clear priorities, and an initialactivation whose impact we can genuinely measure. That’s how AIrepresentation moves beyond a visibility exercise and becomes abusiness lever that can be actively managed. — Giuliano Torelli , CMO at Redify

These are the patterns worth designing against from day 1.

• Measuring once. A single reading of a probabilistic system is an anecdote. Any programme built on it will be argued with, correctly.

• Measuring one engine. Checking ChatGPT and concluding the brand is fine describes that engine’s reading habits, not the brand’s representation.

• Working only on the owned site. 28.3% of ChatGPT’s most-cited pages have no organic Google visibility at all. Ahrefs found 36.7% of AI Overview citations come from domains outside the organic top 100. Search position is a poor proxy for AI citation.

• Starting with the slowest lever. Earned media is the right long-term programme and the wrong first move, because it cannot produce a readable result within 90 days.

Turning a quarter into a system

It is not possible to run all by hand every month. Search Bridge AI Representation Intelligence exists to make the sequence above repeatable.

It works across three tracking layers, correlated into one picture.

Deep Tracking goes inside the AI’s knowledge graph to establish how the brand is understood by AI and which sources impact it the most.

Competitive Tracking establishes where the brand stands against competitors on the real questions buyers ask, across every major engine, and reports share of voice.

Technical Tracking establishes whether AI systems can access, read, and understand the brand’s content: discoverability, navigability, content clarity, schema, llm.txt, and metadata.

What comes out is a set of prioritised Activations, ranked by expected impact and tied to the specific source with the most to gain. That is the plan the first 90 days need, produced continuously and with AI Experts team at hand.

‍

Sources: SparkToro (2026), Search Engine Land (2026), Conductor (2026), Ahrefs (2026), McKinsey (2026), Search in Italy 2026.