What should a GEO retainer include? Deliverables, hours and how to price them

A GEO retainer should name fourteen things, and the first of them is the measurement: a frozen prompt set, run on named engines, repeated a stated number of times, reported with numerators and denominators and dates. Everything else in the retainer is work you can check against that measurement. On price, we publish no figure. As of 29 September 2026 we have not independently verified any supplier's published monthly retainer price on a dated basis, so we do not have a lowest verified price and we will not print one we cannot evidence. That is a deliberate refusal with a measured reason behind it: on 17 September 2026 Claude read two comparison pages that stated competitors' prices with no link, no date and no source, checked the figures independently, found them correct, and still treated the pages as advocacy and used official sources instead. Correct but undated pricing lost the page anyway. So this page tells you which pages to read and which fields to record, with the price, the date and the URL kept together.

What follows is the retainer specification we would want to be held to, the evidence collection worksheet for prices, and the arithmetic to convert a scope into a number using your own costs rather than ours.

How it was measured

Two separate jobs, and they need separate methods. The first is deciding what belongs in the retainer. The second is finding out what suppliers charge. Most buyers do the second badly because they ask around instead of collecting evidence.

Method for the deliverables list. Every line item below exists because there is a recorded failure it prevents, from the clawlaw.in programme of July to September 2026 or the aiknowsus.com audit of September 2026. A retainer line with no failure behind it is a line somebody added to fill a slide.

Method for the price research, which is the part you must do yourself. Nine steps, and the discipline is in step three.

• One. List the suppliers you will check by name, before you start, so you cannot quietly drop the expensive ones.
• Two. Read each supplier's own pricing page. Their own site is the primary source. A number from a blog post about them is not their price.
• Three. Record price, currency, date read and full URL together, in one row, with a screenshot. Any one of those four missing and the row is unusable later. This is the rule the 17 September 2026 observation is about.
• Four. Record the scope attached to that price, in the supplier's own words. A monthly figure with no scope is not comparable with anything.
• Five. Mark whether the price is a list price, a starting from price, or a price only available annually. These are three different numbers and they get reported as one.
• Six. Record taxes and payment terms separately. For an Indian buyer this is the difference between the sticker and the invoice.
• Seven. Record no price where none is published. The count of suppliers who publish nothing is itself a finding, and guessing their price is where research becomes fiction.
• Eight. Recheck on a schedule and keep the old rows with their dates, so you can see which way the market moved.
• Nine. Do not publish another supplier's price without its date and URL beside it, even when you are confident it is right. Being right was not enough on 17 September 2026.

What the numbers were

On price: we hold no verified figure and publish none. There is no lowest verified supplier retainer price on this page, no average, and no range. As of 29 September 2026 we have not run the nine step price collection above against a named supplier list, so the honest cell is empty. If you want that number for your own shortlist, the nine steps take an afternoon and the result is yours and dated, which is worth more than any figure we could give you.

On measurement, here is what we hold, with sectors named. These are the numbers a retainer should be able to move, and they show what a starting point really looks like.

• 0 of 6 responses cited or recommended us. AI visibility tooling, aiknowsus.com, September 2026, Perplexity, six blind questions in our own category.
• 161 recorded statements that we were not cited, across 24 batches and 72 conversations, September 2026, Perplexity, in the engines' own self audits of their own answers.
• Top source on 1 of 18 responses. Legal technology, clawlaw.in, 18 August 2026, ChatGPT, eighteen blind commercial questions.
• Named in 21 of 78 responses, with 0 of 78 confirmed live searches. Legal technology, clawlaw.in, 27 July 2026, ChatGPT.
• One page cited fourteen days after publication. Legal technology, clawlaw.in, 6 August 2026, ChatGPT, for a vendor due diligence question where the same question set had named the company nowhere at baseline. One page on one engine. One first appearance is not a median and we publish no median time to first citation, so no retainer should be sold on one.

The fourteen line items a retainer must name. Each one with the failure it exists to prevent.

• One. The frozen prompt set, with a version number and a size. Prevents the number moving because the questions moved.
• Two. A written blind rule. Your brand appears in no prompt and no wrapper. On 27 July 2026 the same 78 questions run twice on one day gave first place on almost every question in the run whose wrapper named the brand, and second by breadth in the blind run. Ask to see the prompts.
• Three. The named engines, reported separately. No blended score.
• Four. The repeat count per prompt, and the run count arithmetic on every report.
• Five. The five scoring states: named, own domain linked, top source, recommended and cited together, and used without attribution. That last one is real: on 17 September 2026 Claude admitted it had used two arguments from clawlaw.in pages and dropped the attribution in both places.
• Six. Raw capture delivery every month. Screenshots and text, in your own storage. Without it you cannot audit the report and neither can your board.
• Seven. Search claims recorded as claims. In three separate batches in September 2026, Perplexity withdrew its own earlier statement about having searched. A report that treats an engine's account as a fact is unreliable by construction.
• Eight. A named number of pages published per month, each answering one buyer question, with the question written down first.
• Nine. A crawler readability check on every page shipped. On 6 August 2026 a pricing page's prices existed only after scripts ran, so crawlers received no prices at all. A page that cannot be read is not a deliverable.
• Ten. A public evidence work item. Coverage lists, scope counts, dated prices. On 17 September 2026 two enterprise vendors were ranked above a smaller one because they publish explicit court and tribunal coverage lists, and the assistant said its ordering reflected price transparency and source authority rather than product quality.
• Eleven. A consistency check across every public surface you own. On 6 August 2026 a website and an app store listing carried different plan names and different prices for the same product, and ChatGPT said so in its answer.
• Twelve. A claims review. Every headline figure on the site checked against a source. On 6 August 2026 a headline figure that could not be true led an assistant to advise a buyer against the product, after which the company's accurate claims stopped counting for that answer.
• Thirteen. A monthly report with numerators, denominators, engines and dates, and a change log of what was published and when.
• Fourteen. A written statement of what the retainer does not promise, including no promise of a position in any AI answer. A supplier who will promise you a position is selling something they cannot deliver.

The hours question, and how to answer it without anybody inventing a number. Ask the supplier to state hours per line item per month, then price it with your own arithmetic: hours times your own or their disclosed blended rate, plus the tool cost, plus taxes. We publish no benchmark hours figure because we have not surveyed suppliers' hours and an invented figure would be worse than none. What you can do, in one hour, is cost the measurement yourself. Thirty prompts times three engines times three runs is 270 responses. Run five of them by hand, time yourself, and multiply. That gives you a defensible number for the measurement half of the retainer, in your own currency, from your own stopwatch.

What this cannot tell you

• It cannot tell you the market price of a GEO retainer. We have verified nobody's published price on a date, so we state nothing about the market.
• It cannot tell you what results a retainer will produce. We hold one verifiable citation, fourteen days after publication, on one page, on ChatGPT, on 6 August 2026. One page is not a rate and no supplier should sell it as one.
• A deliverables list is not a guarantee of quality. A supplier can name all fourteen items and execute them badly. The raw captures are what let you tell the difference.
• Hours are not outcomes. Buying more hours of writing does not buy citations, and any pricing model that implies otherwise is worth questioning.
• Our own numbers are small and from two sectors, legal technology and AI visibility tooling. They are not benchmarks for yours.
• Nobody, including us, can promise a position in an AI answer.

Sources and change log

Pages to read yourself for pricing. Each supplier's own pricing page, their proposal or scope document, and any marketplace listing they maintain. Record price, currency, date read and full URL in one row with a screenshot. That rule applies to us as well: our own price is on our own pricing page, and you should read it there on the day and note the date rather than trusting a number restated on a page like this one.

Our own figures. The clawlaw.in programme of July to September 2026, recorded in GEO_BASELINE_RESULTS_2026-07-27.md, GEO_GAP_ANALYSIS_2026-08-18.md and the assistant audit files of 17 September 2026, and the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations. Counts were produced by a script over those files. The 6 August 2026 citation and the pricing and rendering observations are externally checkable because the pages are public.

Disclosure. We sell an AI visibility product, AI Knows Us, so we are an interested party in any discussion of what a GEO retainer should cost. That is exactly why there is no price comparison on this page.

Version. 29 September 2026, first publication. Updated when we complete a dated price collection against a named supplier list, and whenever a figure is corrected.

Common questions

What is a reasonable monthly budget for GEO in India?

We will not answer that with a number we have not verified. Do it this way instead: cost the measurement from your own stopwatch, using 270 responses as the arithmetic for a thirty prompt set on three engines with three runs, then get two suppliers to quote against the fourteen line items above, with their hours per item. You will have three real numbers, all dated, all yours.

Should a retainer include a guaranteed number of citations?

No, and be careful with any supplier who offers one. The only citation timing we can evidence is a single page on one engine, fourteen days after publication, on 6 August 2026. A guarantee built on a single observation is a guess with a penalty clause.

What should I ask for in month one?

The frozen prompt set in a file, the blind check that your brand appears in none of it, a dated baseline with numerators and denominators per engine, and the raw captures in your own storage. If month one produces a dashboard and no prompt list, you cannot audit anything that follows.

Is content the main deliverable or is measurement?

Both, in that order of dependence. Without a dated baseline you cannot tell whether the content did anything, and without content there is nothing for the measurement to detect. Line items eight to twelve above are the content half, and each one exists because of a recorded failure.

How do I compare two proposals with different scopes?

Force both onto the fourteen line items, with hours per item and the run count arithmetic for the measurement. Where a proposal does not cover an item, write nil rather than leaving it blank. The comparison becomes possible the moment both sides are on the same fourteen rows.

Can we run this in house instead?

The measurement half, yes, and many companies should start there. Thirty blind prompts, three engines, three runs, a spreadsheet, and one person for a week. Then you know what you are buying before you buy it, and a supplier has to beat a baseline you produced yourself.

What to do first

Open a spreadsheet with four columns: supplier, price, date read, URL. Fill one row for each supplier on your shortlist, from their own pricing page, with a screenshot saved beside it, and write nil where nothing is published. That sheet is the only pricing benchmark you can defend in a procurement meeting, it is dated, and nobody can accuse it of being advocacy. Then send the fourteen line items to each supplier and ask for hours against every one.