How to Do Generative Engine Optimization in 90 Days: Audit, Fix, Measure

Quick answer

Canlah AI does generative engine optimization in four phases across 90 days under the method in this guide: days 1 to 15 freeze a buyer-question panel and record a baseline with no optimisation, days 16 to 45 repair entity and structured-data hygiene, days 46 to 75 build off-site authority and days 76 to 90 re-run the identical panel and report intervals instead of point values. The order is the point, because on-site editing, the phase most programmes run first, has the weakest evidence behind it.

The sequence is published in Chapter 7 of Canlah AI’s 2026 whitepaper and was run on canlah.ai itself. The self-audit of September 7, 2026 shows why the order matters: the site was named in 39 of 39 grounded branded answers and in 7 of 177 grounded open answers, a 4.0% non-branded mention rate, while the pipeline’s composite score read 71 out of 100 and its GEO dimension read 4 out of 100.

Treat one 90-day result as directional. A critical review of 45 GEO studies dated July 15, 2026 found no technique in its scope with a stable, longitudinal, cross-platform causal effect. Build the programme to produce evidence somebody else can re-run, not a lift you have to be trusted on.

Disclosure: Canlah AI publishes this guide and sells the measurement and execution service the sequence describes. The first-party figures come from a self-audit of canlah.ai run on September 7, 2026 and from Canlah AI’s 2026 whitepaper. Tool prices and plan limits were read from public pricing pages on September 23, 2026 and were not confirmed through paid accounts.

What does a generative engine optimization programme change?

A generative engine optimization programme changes six measurable things, and only the first two are visible without instrumentation. The definition of the discipline sits in the GEO and SEO comparison; this page is the implementation sequence.

• Mention: Whether an answer names the brand on questions that do not contain the brand name.
• Recommendation position: Whether the brand is the main pick, one option in a shortlist or a passing reference.
• Citation: Whether the answer links to the brand’s page, a third-party page, a competitor or no visible source.
• Source mix: Which domains supply the evidence, and whether the brand appears on any of them.
• Accuracy: Whether category, market, product, price and capability are described correctly.
• Engine: Whether the result holds across ChatGPT, Gemini, Google AI Overviews and Google AI Mode, or only on one of them.

The last three decide what work to commission. A single visibility score with no cited-domain list underneath cannot tell you which phase below deserves the quarter.

Why a GEO framework starts with measurement, not content

A GEO framework starts with measurement because Canlah AI’s September 7, 2026 self-audit found canlah.ai named in 39 of 39 branded answers and in only 7 of 177 open answers, a gap no on-site edit was positioned to close. In the evidence Canlah AI reviewed, on-site scores did not correlate with discovery rate, while referring domains and third-party presence did. On-site work runs first in most programmes because it is the easiest thing to deliver, not because it is the strongest lever.

The self-audit is the cleanest illustration available because it is the one audit with no client to protect. Across 39 questions in six funnel layers, three runs each, the site answered every question about itself and almost none that precede it. The engines disagreed outright: ChatGPT, on the OpenAI API with web search, produced all seven open mentions, 7 of 84 answers or 8.3%, while Gemini produced 0 of 93.

The rendered Google leg said the same thing. An AI Overview appeared on 62 of 65 valid Singapore renders, named Canlah AI on 11 of 11 branded renders and on 0 of 51 non-branded renders. Meanwhile 381 third-party domains were cited for those questions, and canlah.ai ranked 46th of 382 cited domains on open-question citations alone. No amount of homepage editing moves a number like that, which is why editing is not phase one.

The 90-day GEO roadmap at a glance

The 90-day GEO roadmap has four phases: baseline on days 1 to 15, hygiene on days 16 to 45, off-site authority on days 46 to 75 and re-measurement on days 76 to 90. Each phase of this generative engine optimization strategy is defined by the record it leaves behind rather than the work it feels like, which is what makes the quarter auditable by somebody other than the vendor.

The 90-day generative engine optimization roadmap Canlah AI publishes and runs, reviewed September 23, 2026.

Phase	Days	What you do	What to record	Decision it supports
1. Baseline	1 to 15	Pre-register a buyer-question panel and run it across engines, with no optimisation	Panel, generation rules, engine, market, language, date, run count, timestamped captures	Whether the problem is recall, accuracy or position, and on which engine
2. Hygiene	16 to 45	Entity consistency, product structured data, review and Q and A markup, platform preconditions	Name, category and market string used in each place; schema types present per template	Whether an engine can tell this brand from similarly named ones
3. Off-site authority	46 to 75	Trade media, professional communities and original data, plus two or three vertical platforms in Chinese markets	Cited-domain list per engine with counts and dates, plus every placement with its live URL	Which outreach targets are worth the next quarter
4. Re-measurement	76 to 90	Re-run the identical panel and split attributable change from sampling variance	Mention-rate interval per engine, source-set change, each result marked attributed or not	Whether to repeat, re-scope or stop

Source: Canlah AI whitepaper Chapter 7 (published August 2026) and the September 7, 2026 self-audit run on the same sequence. Phase lengths are an operating default, not a benchmark.

Days 1 to 15: freeze the panel and record the baseline

Phase one, days 1 to 15, produces two things only: a frozen panel of 30 to 60 buyer questions and a baseline mention rate per engine. Skipping it is the most expensive decision in the programme, because every later claim of improvement needs a before recorded under the same conditions.

1. Write 30 to 60 buyer questions across the funnel, from category awareness to shortlist to brand verification, drafted by somebody who does not know which answer flatters you.
2. Publish the panel and its generation rules before the first run, so the criteria cannot move once results arrive.
3. Keep branded and open questions in separate layers and never average them. In Canlah AI’s audit the two layers read 100.0% and 4.0%.
4. Run each question at least three times per engine, and state that the source set is a sample, not an enumeration.
5. Record engine, model version, market, language, account state, interface and timestamp on every run, with browser-grade captures for anything you will quote.
6. Log factually incorrect statements separately from the visibility count. Accuracy failures are repaired differently from absence, and carry legal exposure that absence does not.

Days 16 to 45: entity and structured data hygiene

Phase two, days 16 to 45, repairs entity and structured-data hygiene, and the gaps are usually wide: Canlah AI’s whitepaper found product structured data missing on 81% of the brands audited and review or Q and A markup missing in over 90% of cases. The phase is hygiene rather than leverage, but every mention phase three earns depends on it.

1. State one name, one category, one market and one description identically on the site, Wikidata, LinkedIn, Crunchbase and the Google Business Profile. Ambiguity misattributes every mention phase three earns.
2. Publish a crawlable facts page an engine can quote without inference, carrying the exact figures, the service scope and the dates.
3. Add product, review and Q and A structured data to the templates that carry commercial answers, then validate them rather than trusting the plugin.
4. Verify the platform preconditions your category depends on. In Canlah AI’s open dataset of 50 cross-border DTC brands, 47 reachable, 0 of 47 exposed an agent-payments endpoint and 1 of 47 an MCP endpoint.
5. Write one exact first-party figure and one market-range sentence on each commercial page. ChatGPT tends to cite the source page for the exact number; Google AI Mode tends to cite secondary syntheses for the range.

Days 46 to 75: off-site authority, the principal lever

Phase three, days 46 to 75, is where the open-question mention rate moves, because referring domains and third-party mentions are the operable variables with the strongest observed correlation to discovery. It is also the phase most buyers are never sold, because it is slower to deliver than a content calendar.

1. Read the cited-domain list from phase one and treat it as the target list. Canlah AI’s audit recorded 381 third-party domains cited on questions where the brand was absent; those are the pages an engine already reads.
2. Prioritise trade media, professional communities and review platforms that carry your category in your market, not the global domains a generic ranking names.
3. Publish original data rather than commentary. A dataset with a method and a re-verification script is quotable; an opinion post is not.
4. In Chinese-language markets, select the two or three vertical authority platforms tied to the ecosystem behind each target model, not the Western source map.
5. Do no volume seeding and operate no account matrices. Canlah AI publishes that red line because the boundary between content optimisation and undetectable manipulation cannot be drawn technically, only declared and then checked against conduct.

Days 76 to 90: re-measure and attribute honestly

Phase four, days 76 to 90, re-runs the identical panel and reports intervals rather than point values, the only step that turns 75 days of work into evidence. Changing the panel here destroys the comparison phase one paid for.

1. Run the same questions the same number of times on the same engines, then record the model version again. Model updates are a competing explanation for any movement.
2. Report mention rate as a range per engine with the run count attached, never as one number for all engines.
3. Separate change in mention rate, which can be attributed to an action, from change in the source set, which is usually sampling variance.
4. Label any change that cannot be placed in a time correspondence with a specific action as unattributed, and leave it that way in the client report.
5. Keep the branded layer as a control. If branded and open rates move together, suspect the instrument before crediting the work.

Generative engine optimization best practices that survive the evidence

Five generative engine optimization best practices held up across the sources reviewed for this guide, and each is a constraint rather than a tactic.

1. Publish the panel before the first run, because an unregistered panel can be edited into a result.
2. Report ranges with run counts and dates, because a composite score hid Canlah AI’s own 71 against 4 out of 100 split.
3. Treat each engine as a separate market, because ChatGPT and Gemini returned seven and zero open mentions on the same panel.
4. Correct inaccurate statements before chasing exposure, then spend the middle of the quarter off-site.
5. Write “association, not causation” wherever a before-and-after figure appears.

How to move an existing SEO programme to GEO

An SEO to GEO transition keeps six existing SEO assets and replaces the measurement layer around them, so it is an extension rather than a migration. Clicks remain the outcome for part of the funnel, but they are no longer the only evidence the brand was considered: a SparkToro and Similarweb panel cited in Canlah AI’s whitepaper put the share of US Google searches ending without any click at 68.01% for January to April 2026, up from 60.45% in 2024.

What an existing SEO programme keeps, changes and adds in the 90 days, mapped September 23, 2026.

What you already have	What it becomes in GEO	What to add in the 90 days	Why it matters
Keyword list	Buyer-question panel, phrased the way a person types into an assistant	Branded and open layers kept separate, three runs per question	Averaging the layers would have hidden a 100.0% against 4.0% split in Canlah AI’s self-audit
Rank tracker	Mention, position and citation rates per engine, reported as ranges	A second and third engine; a single-engine read is not a programme	ChatGPT returned 7 of 84 open mentions and Gemini 0 of 93 on the same panel
Backlink targets	Cited-domain list read from your own baseline	Placement records with live URLs and dates, per engine	381 third-party domains were cited on questions where canlah.ai was absent
Technical audit	Entity resolution plus product, review and Q and A structured data	A crawlable facts page with exact figures an engine can quote	Product structured data was missing on 81% of brands in the whitepaper audit
Content calendar	One exact figure and one market-range sentence per commercial page	Original data with a published method and re-verification script	ChatGPT tends to cite the exact figure; Google AI Mode tends to cite the range
Monthly traffic report	Mention-rate interval and source-set change, attributed or unattributed	Panel, denominators and timestamps inside the report	68.01% of US Google searches ended without a click from January to April 2026

Source: Canlah AI’s 2026 whitepaper and the September 7, 2026 self-audit. “Unattributed” means the change could not be placed in a time correspondence with a specific action under this method; it is not a claim that the action had no effect.

How to interpret four common 90-day outcomes

Four outcomes cover most 90-day GEO re-measurements, and only the first supports crediting a visibility gain to the work without another run.

Mention rate up, source set unchanged: This is the cleanest result available. The same domains were read and the brand now appears among them, which points at phase three placements or the facts page rather than at the model.

Mention rate up, source set changed: Treat this as unattributed until the next run. The engine sampled a different set of pages, so the movement may be variance rather than work.

Mention rate flat, accuracy improved: Phase two worked and phase three has not landed. The engine describes the brand correctly when it reaches it and still does not reach it on open questions, the 39 of 39 against 7 of 177 pattern. Extend off-site work rather than rewriting pages.

One engine moves, the others do not: This is the normal outcome, not a failure. Report it per engine and check whether that engine’s cited domains overlap the placements made. Then decide whether the other engines justify a source map of their own.

What 90 days of generative engine optimization cannot prove

A 90-day generative engine optimization programme cannot prove five things, however clean the re-measurement looks.

• It cannot prove a stable share of voice across every customer question, because the panel is a sample and the answers are stochastic.
• It cannot establish causation from one before-and-after pair, which is what the July 15, 2026 review of 45 GEO studies found across the published literature.
• It cannot guarantee future mentions, citations or recommendations after a model or index update.
• It cannot reproduce a signed-in consumer interface if the measurement ran through an API, so the two surfaces are reported separately.
• It cannot replace factual review, legal approval or analytics attribution.

When to run the second 90 days

Canlah AI recommends a second 90-day cycle in three cases: when the open-question mention rate has moved and you want to know whether it holds, when a new market or language enters the plan or when an inaccurate statement is still circulating. Repetition converts a snapshot into a trend.

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. Its limitation is scope: this sequence is built for brands that can spend 30 of the 90 days on off-site work, and a business that needs a mention within 30 days is better served by paid placement.

Do not take this roadmap, or any vendor’s, on trust. Ask for the panel, the denominators and the dates before the first invoice, then re-run three of the questions yourself on the engines the report names.

Frequently asked questions

How do you do generative engine optimization step by step?

Generative engine optimization is done in four phases across 90 days: freeze a buyer-question panel and record a baseline on days 1 to 15, repair entity and structured-data hygiene on days 16 to 45, build off-site authority on days 46 to 75 and re-run the identical panel on days 76 to 90. Canlah AI publishes this sequence in Chapter 7 of its 2026 whitepaper. The order matters because on-site editing, the step most programmes start with, has the weakest evidence behind it.

How long does generative engine optimization take to show results?

Generative engine optimization takes 90 days to reach a defensible re-measurement under the published sequence, with days 1 to 15 on the baseline alone. Entity corrections can surface within weeks because engines re-read profile pages often. Off-site authority accumulates over a quarter, which is why re-measurement sits at day 76, not day 30.

Can I do generative engine optimization myself without an agency?

Yes, you can run the measurement half of generative engine optimization yourself with a self-serve tracker, which covers phases one and four. OtterlyAI lists Lite at US$29 a month for 15 prompts, Standard at US$189 for 100 and Premium at US$489 for 400; Peec AI’s Starter plan covers 50 prompts across three models; Profound offers a free trial of 50 prompts daily for seven days across ChatGPT, Gemini and Google AI Overviews. All three pricing pages were read on September 23, 2026. Phases two and three are execution work that no dashboard performs.

Is generative engine optimization just SEO with a new name?

No. Generative engine optimization shares the SEO asset base but not its measurement layer. SEO reports position and clicks on a deterministic index; GEO reports mention, position, citation and accuracy across non-deterministic engines that cite different source graphs, which is why Canlah AI’s self-audit saw the same questions produce 7 open mentions on ChatGPT and 0 on Gemini.

What should a 90-day GEO programme cost?

A 90-day GEO programme costs US$29 to US$489 a month for published self-serve tracking, while agency retainers start near US$1,500 and run past US$25,000 for enterprise scope, based on pricing pages reviewed September 23, 2026. Canlah AI starts with a free AI visibility report within 48 hours, and its GEO service is priced per AI platform: Starter from $1,500 a month (ChatGPT, 4 buyer questions), Growth from $3,000 a month (adds Google AI Overviews, 8 buyer questions) and Full customised (adds Google AI Mode and Gemini, 16 buyer questions). Price the phases separately: phase two is bounded work and phase three is not.

How many questions and engines does a credible baseline need?

A credible generative engine optimization baseline needs enough questions and runs that every reported rate has a denominator you can see. Canlah AI’s audit used 39 questions in six layers, three runs each, across an OpenAI leg, a Gemini leg and a browser-rendered Google AI Overviews leg, and still called its source set a sample. Two engines is the minimum for a cross-engine claim; one engine is a single-vendor observation.

Method and sources

This guide was written on September 23, 2026. The sequence, the phase boundaries and the red line come from Chapter 7 of Canlah AI’s 2026 whitepaper. The first-party figures come from the September 7, 2026 self-audit of canlah.ai: 39 English questions in six funnel layers, three runs each, on an OpenAI leg, a Gemini leg and a browser-rendered Google AI Overviews leg for the Singapore market. That archive is one agency measuring one site in one market on one day, not a randomised experiment, and the audit’s own usability review judged the report unfit to send to a client in that state. Tool prices were read from public pricing pages on the date stated, not through paid accounts.

• Canlah AI whitepaper, Chapter 7. The 90-day sequence, the phase boundaries and the published red line.
• Canlah AI whitepaper, Chapter 3. The July 15, 2026 critical review of 45 GEO studies.
• Canlah AI GEO audit of its own site. Mention counts, engine split, cited-domain totals and the composite-score finding.
• Canlah AI AI Info. The measurement protocol used in phases one and four.
• Agent-Readiness of 50 Cross-Border DTC Brands. The agent-payments and MCP endpoint counts quoted in phase two, CC BY 4.0.
• OtterlyAI pricing. Lite, Standard, Premium prompt limits and prices, reviewed September 23, 2026.
• Profound pricing. Free trial scope and enterprise engine count, reviewed September 23, 2026.
• Peec AI pricing. Starter, Pro and Advanced prompt and project limits, reviewed September 23, 2026.

See where your brand stands in AI answers. Run the free AI visibility audit →

Related articles

• GEO vs SEO: what is the difference
• Generative engine optimization cost in 2026
• How to improve AI search visibility for your brand
• We ran our own GEO audit
• The GEO vendor evidence checklist