How to Rank in ChatGPT: The Five Gates to Getting Cited

Updated September 2026 · Written and maintained by the Progression Agency strategy team

There is no ranked list inside a ChatGPT answer. There are five gates between your page and a citation, and most sites fail at the first two — where the fix costs hours, not quarters. This is the whole method.

How do you rank in ChatGPT?

You do not rank in ChatGPT, because ChatGPT has no ranked list. You become one of a handful of sources it retrieves and quotes, and you get there by passing five gates in order: the crawler can fetch you, the answer exists in the HTML, the page is indexed, the passage is retrieved for the question, and the passage is clear enough to quote.

The word ranking is doing a lot of damage here. It implies a list with positions, a scoreboard you can climb by doing more of something. ChatGPT does not produce a list. It produces one answer, assembled from whatever it retrieved for that particular phrasing on that particular run, and then it attaches a small number of links. There is no second page. There is no position four. You are either in that handful of sources or you are not.

That distinction changes what the work looks like. Ranking work is comparative and incremental: you try to be slightly better than the page above you. Citation work is binary at every gate and then probabilistic at the end. Four of the five gates are pass or fail, and most of the sites we audit are failing one of the first two, where the fix costs hours rather than quarters.

This page is the full method: what each gate is, how to tell whether you are passing it, what to do when you are not, what does not work despite being widely recommended, and how to measure the result honestly. It is the same sequence we run as a paid engagement. Nothing important is held back, because the difficulty here is execution and measurement discipline, not secrecy.

What actually decides whether ChatGPT cites you?

Retrievability first, then clarity, then corroboration. Whether you are reachable, whether your answer is extractable, and whether anyone other than you says the same thing.

Retrievability is the part nobody wants to talk about because it is unglamorous. It is robots files, server responses and whether your content survives without JavaScript. It is also where the overwhelming majority of failures live. A site that cannot be fetched is not competing badly; it is not competing.

Clarity is the editorial half. A passage gets quoted when it answers a question in a form that can be lifted out of context and still make sense. Long windups, hedged qualifications and sentences that only work if you have read the previous three paragraphs are the enemy of extraction. This is not about writing simply; it is about writing so that a single paragraph carries a complete, attributable claim.

Corroboration is the slow half. Assistants weigh what other sources say about you. A claim that appears only on your own site is an assertion; the same claim appearing in trade press, a directory, a review platform and a client’s own write-up is a fact with several witnesses. This is the part that cannot be bought quickly and the part that most reliably separates the businesses that get cited from the ones that do not.

It is not keyword matching

Retrieval works on meaning rather than exact terms, so the old instinct to repeat a phrase until it appears often enough does nothing useful here. What matters is whether a passage is clearly about the thing being asked. A page that says the same thing six ways is not six times more retrievable; it is one retrievable passage surrounded by noise.

It is not a submission process

There is no queue to join, no form that registers your site with OpenAI, and no verification badge that changes your standing. Anyone selling submission is selling nothing. The only actions that affect your standing are the ones on your own site and the ones other people take on theirs.

It is not purchasable

There is no advertising slot inside a ChatGPT answer that you can buy today, and no paid tier that places you in one. That is worth stating plainly because the absence of a paid shortcut is exactly why the technical and editorial work has leverage: everyone competes on the same terms, and most competitors have not done the work.

It is not stable between runs

Ask the same question twice and you can get two different sets of sources. This is not a fault in your setup; it is how the systems behave. It means any single observation is worthless as evidence, and it is the single strongest argument for measuring against a fixed prompt set repeated over time rather than reacting to whatever you saw this morning.

Which ChatGPT surface are you actually trying to reach?

Two of them, and they behave differently. The model’s own knowledge is slow, broad and rarely linked. The search surface is live, narrow and it links. Almost all of the work on this page targets the search surface, because that is the one you can influence on a useful timescale.

When ChatGPT answers from what the model already holds, your site is not being fetched at all. That knowledge was assembled long before the question was asked, changes only when the model does, and usually produces an answer with no links in it. You can influence it, but the lever is the general presence of your business across the open web over a long period — not anything you ship this quarter.

When ChatGPT searches, it fetches pages in response to the question, and it attaches links to what it used. This surface is where a technical fix made on Tuesday can matter by the following month. It is also the surface that sends actual clicks, because it is the one that shows links. When a client says they want to rank in ChatGPT, this is nearly always what they mean, even if they have not separated the two.

The model-knowledge surface

Broad reach, no freshness, few links. Influenced by how widely and consistently your business is described across the web over years. There is no fast lever here, and anyone promising one is describing the other surface without saying so.

The search surface

Narrow reach per answer, high freshness, links attached. Influenced by crawler access, rendering, indexation and passage clarity — all of which are things you control directly. This is where the five gates apply and where measurable movement happens.

Why the distinction matters commercially

A programme aimed at the model-knowledge surface is a multi-year brand exercise. A programme aimed at the search surface is a technical and editorial project with a defined scope. They cost different amounts, take different lengths of time, and are frequently sold as the same thing. Ask any provider which one they are proposing.

Step one: let the crawlers in, and prove it

Allow GPTBot, OAI-SearchBot and ChatGPT-User in robots.txt, confirm your WAF or bot management is not blocking them separately, then read ninety days of server logs to prove they arrived. Until the logs show a fetch, nothing else you do is reaching anything.

This is the cheapest and most commonly failed step. Two things block crawlers and they are usually managed by different people. The first is robots.txt, which is often edited once during a site launch and never revisited; a blanket disallow added to keep a staging site quiet has a way of surviving into production. The second is the bot-management layer in a CDN or WAF, which blocks unfamiliar user agents by policy and does so invisibly, because robots.txt looks fine and nobody thinks to check the other system.

The proof is not the robots file. The proof is the log. A robots file states an intention; a server log records what actually happened. Pull ninety days, filter for the user agents by name, and count the fetches. If the count is zero for a crawler you believe you have allowed, something between your intention and the internet is intercepting it, and finding out what is the single highest-value hour in this entire process.

Check robots.txt, in production, today

Read the live file rather than the one in your repository. They diverge more often than anyone expects, particularly on sites where a platform generates the file dynamically or a plugin appends rules you did not write.

Check the bot-management layer separately

Request your own pages with each crawler’s user-agent string and record the status code. A 403 returned only to bots is invisible in every browser test, and it is the failure mode that survives longest because nothing on your side looks broken.

Check rate limiting

Some sites allow crawlers and then throttle them so aggressively that a full crawl never completes. The logs show fetches, so it looks like a pass, but the pages that matter are never reached. Look at which URLs were fetched, not only how many.

Decide deliberately about content rights

Allowing crawlers means your content can be used in answers people read without visiting you. That is a real trade-off and some publishers reasonably decline it. What is not reasonable is making that decision by accident, which is what a stale robots file amounts to.

Step two: make the answer exist without JavaScript

Fetch your own pages with JavaScript disabled and read what comes back. If the answer is not there, no amount of content work will help, because the crawler is reading the same empty shell you are looking at.

This is the expensive failure, and it is the one that looks perfect in a browser. Modern frameworks assemble pages in the visitor’s browser. The HTML that leaves your server can be a few hundred bytes of scaffolding plus a script tag, with every word of the actual content arriving afterwards. Humans never notice because their browser runs the script. Many crawlers execute little or no JavaScript, so they receive the scaffolding and nothing else.

We have measured pages that ranked for their target term and returned under two hundred words when fetched without JavaScript. Every content improvement made to those pages was invisible to the systems the business was trying to reach. The team producing that content had no way of knowing, because every tool they used rendered the page the way a browser does.

The fix is server-side rendering or pre-rendering for the templates that carry commercial content. On most stacks this is a configuration change rather than a rebuild, and it does not require abandoning your framework. The work is bounded: identify the templates that matter, render those server-side, verify, move on.

How to test it honestly

Fetch the URL with a plain HTTP client and read the body, or disable JavaScript entirely in a fresh browser profile and reload. Do not use a rendering-based crawler for this test; its whole purpose is to execute the script you are trying to test without.

Test templates, not pages

Failures cluster by template. If one service page is empty without JavaScript, every service page is. Testing twenty individual URLs from the same template tells you nothing more than testing one, and testing one from each template tells you everything.

Watch for partial rendering

Some setups render the page shell and the navigation server-side while loading the body content client-side. That is the worst case, because automated checks see a populated page and report success while the content the crawler needs is still missing.

Check the response time while you are there

Retrieval has tighter timeouts than indexing. A page that takes four seconds may simply not be fetched, and a slow page is a silent failure rather than a visible one. Measure from a bot’s perspective, not from a warm browser cache on your office connection.

Step three: put the answer first, in every section

Write the heading as the question someone would actually ask, then answer it in the first sentence or two, then explain. This is the single editorial change that moves the most, and it costs one pass over your twenty most commercially important pages.

Retrieval operates on passages, not whole pages. A system looking for an answer to a specific question is looking for a chunk of text that contains one. If your section opens with three paragraphs of context and reaches the answer in the fourth, the chunk that gets evaluated is the context, and the context does not answer anything.

This is a change in structure rather than in substance. Nothing has to be simplified, shortened or dumbed down; the explanation still follows in full. What changes is the order. Lead with the claim, then justify it. Journalists have written this way for a century for a related reason: so that a reader who stops early still has the point.

There is a useful side effect. Pages restructured this way perform better with human readers too, particularly the ones arriving from a search with a specific question. The change is not a concession to machines; it is a concession to anyone in a hurry.

1. Turn each H2 into the question a buyer would type or say.
2. Answer it in the first sentence after the heading, in one or two sentences.
3. Put the reasoning, caveats and detail after that, in full.
4. Keep each section to one idea, so the passage is self-contained.
5. Name the source of any claim inside the same paragraph as the claim.
6. Remove the transitional sentences that exist only to introduce the next heading.
7. Cut hedging that does not change the meaning; keep hedging that does.
8. Re-read each section alone and ask whether it still makes sense out of context.
9. Fix the ones that do not, because those are the ones that will never be quoted.
10. Do this for twenty pages before doing it for two hundred.

Why the first 300 characters matter

That is roughly the length of the fragment an assistant will lift. If your answer is not complete within it, what gets lifted is an incomplete version of your point, and an incomplete point is less likely to be used at all.

Why one idea per section

A section covering three related ideas produces a passage that answers none of them cleanly. Splitting it into three sections with three question headings produces three retrievable passages. This is the rare case where more structure genuinely produces more surface area.

Why hedging hurts

A sentence that qualifies itself into ambiguity cannot be quoted as an answer, because it does not contain one. Keep the qualifications that carry real information — timescales, conditions, exceptions — and cut the ones that are only defensive.

What not to do

Do not add a question-and-answer block at the bottom of a page that contradicts the page above it, and do not convert your entire site into FAQ format. Structure helps when it reflects the content. It hurts when it is a costume.

Step four: make every claim checkable

Attribute claims to named, linkable sources inside the paragraph that makes them. Unsourced confidence reads as marketing copy, and marketing copy is the category of text these systems are most careful about reusing.

There is a practical asymmetry here. A sentence that says a rule exists is a claim. A sentence that says a rule exists and links to the body that published it is a claim with a witness. The second is safer to quote, and safety is a large part of what governs whether a passage gets used in an answer that someone might act on.

This also protects you from a real risk. Superlatives that cannot be verified — largest, fastest, highest, only — are exactly the sentences that get repeated back to a prospect who then checks them. We have had to correct a published superlative on our own work after verifying it against primary data and finding it wrong. Checkability is not a stylistic preference; it is a way of not being caught out.

Link to primary sources, not summaries

A regulator’s own page, a standards body’s own document, a platform’s own documentation. Secondary coverage of a primary source adds a link and subtracts confidence.

Date your figures

A number without a date is unusable to anyone checking it, and it ages invisibly. A number with a date and a source can be re-verified by anyone, including you, next year.

Say what you do not know

Stating the limits of your evidence is not weakness. It is the difference between a page that reads as informed and one that reads as promotional, and the distinction is visible in the text itself.

Never invent evidence

Fabricated tests, invented client results and imaginary surveys are catastrophic if found, and they are found. If you have not measured something, describe what you have observed and say that is what it is.

Step five: get corroborated somewhere other than your own site

Assistants weigh independent mentions heavily. What you say about yourself is one source. What four unrelated sites say about you is a pattern, and patterns survive scrutiny.

This is the slowest of the five steps and the one with the longest tail. It is also the one that most clearly separates businesses that get cited from businesses that have done everything technical correctly and still are not. Once retrievability is solved and the writing is clear, the remaining difference is usually whether anyone else has ever written about you.

The useful forms of corroboration are unglamorous: trade coverage, accurate directory listings, named case studies hosted by the client rather than by you, conference and association listings, review platforms, and citations in other people’s writing. None of these are purchasable in bulk and all of them compound.

The unhelpful forms are the ones sold in bulk: link packages, syndicated press releases posted to hundreds of identical sites, and directory submissions to aggregators nobody reads. These add repetition without adding witnesses, and repetition of a single source is not corroboration.

Consistency matters more than volume

Four listings that describe your business identically are worth more than forty that describe it four different ways. Inconsistency is the thing that produces wrong answers about who you are, and wrong answers are worse than no answer.

Client-hosted evidence is the strongest

A case study on your client’s own site, with their name on it, is corroboration in a way that the same case study on your site is not. It is also harder to get, which is precisely why it carries weight.

Reviews count, but as voice not as authorship

Review platforms establish that other people describe you, which is the point. They do not establish expertise and they are not a substitute for coverage.

This is the part you cannot rush

Everything else on this page can be done inside a quarter. This one runs for years, which is the argument for starting it now rather than after the technical work is finished.

Step six: fix what your site says you are

Write one canonical description of the business and use it without variation everywhere — your site, your markup, your listings, your profiles. Three different self-descriptions produce three different answers, and you do not get to choose which one an assistant repeats.

Entity confusion is a quiet, common and completely fixable failure. A business describes itself one way in the page copy, a second way in the Organization schema, and a third way on its social profiles, usually because the three were written years apart by different people. An assistant asked what the business does will produce one of the three, and it will often be the oldest, because that one has had the longest to spread.

The fix is administrative rather than technical. Decide the description. Apply it everywhere. Keep a record of where it is applied so it can be updated in one pass when it genuinely changes. This takes an afternoon and prevents a category of wrong answer that no amount of content production will correct.

Step seven: measure it properly, or do not claim it worked

Build a fixed set of prompts before you change anything, record the answers, and re-run the identical wording monthly. Without a baseline, every later claim of improvement is unfalsifiable — including your own.

Two measurement systems are needed and they should never be merged. The technical measures come from logs and rendering: which crawlers arrived, which URLs they fetched, what the page contains without JavaScript, what the response time was. These are certain, fast and binary. They tell you whether the work was done.

The visibility measures come from sampling prompts: how often you appear, whether you are linked, what the assistant says about you, and who else appears. These are slow and probabilistic, because the same prompt produces different answers on different runs. They tell you whether the work mattered.

Reporting these together is how bad programmes hide. A report that shows crawler access improving and calls it visibility is describing an input as though it were an outcome. Keep them in separate sections with separate confidence levels, and say plainly which is which.

Fix the wording, permanently

The phrasing of a prompt changes the answer more than most people expect. If you reword your prompts between runs you have no comparison, only two unrelated observations. Write them down and do not improve them.

Sample enough prompts

A handful is noise. Sixty to two hundred prompts across your real commercial questions gives a rate you can compare month to month. Fewer than that and a normal run-to-run variation looks like a trend.

Record who else appears

The other names in the answer are your actual competitive set for this channel, and they are frequently not the competitors you track in search. This is often the most immediately useful output of the whole exercise.

Distinguish a mention from a citation

Being named in prose is not the same as being linked. Both are worth recording; only one sends traffic. A report that counts them together is overstating the result.

Expect it to be noisy

Month-to-month movement of a few percentage points is normal variation. Treat a trend as real when it persists across three measurements, not when it appears in one.

What does not work, despite being widely recommended?

Paying for placement, submitting your site, treating llms.txt as a strategy, keyword density, and publishing more posts. Five of these are free to stop doing and one of them is expensive.

Every new channel produces a layer of tactics that sound plausible and do nothing, and this one is young enough that the layer is thick. The common thread is that they are all things a marketing team can do without involving a developer, which is exactly why they get recommended and exactly why they do not touch the gates that actually matter.

Paying for placement

There is no ad inventory inside a ChatGPT answer to buy today. Anyone offering to place you in one is either describing something else or selling something that does not exist.

Submitting your site

There is no submission endpoint that affects citation. Crawler access is controlled from your side, not by registering with anyone.

llms.txt as a strategy

Adoption is limited and it is not a ranking factor. It costs ten minutes to add and there is no reason not to, but it belongs in the same category as a well-formed sitemap: hygiene, not leverage.

Keyword density

Retrieval is not term-frequency matching. Repeating a phrase produces a page that is harder to read and no more retrievable.

Publishing more posts

This is the expensive one. Volume produced to hit a cadence dilutes the pages that could have been cited, consumes the budget that should have gone to rendering and structure, and creates a maintenance liability that grows every month.

Buying an AI visibility score

Several tools produce a proprietary number. The number is not wrong exactly, but it is not comparable to anything outside that tool, and it is not evidence. Your own prompt set is.

How do you check whether ChatGPT already cites you?

Ask it the category question rather than your brand name, look for a link rather than a mention, and repeat the run several times. Ten minutes of this is worth more than any tool you could buy this week.

Start with the question a buyer would ask, phrased the way a buyer would phrase it — who does this kind of work in this place, or what should I use for this specific problem. Do not start with your own name, because asking an assistant about a named company almost always produces something, and that something tells you nothing about whether you would ever surface for someone who does not already know you.

Then look at what came back. Were you named? Were you linked? Which other businesses appeared, and are they the ones you expected? Run it again, and again, and watch how much the answer moves. That variance is the most important thing you will learn, because it is the reason a single check proves nothing and a fixed repeated set proves something.

• Ask the category question, not the brand question, first.
• Phrase it three different ways and record all three answers.
• Note whether you were named, linked, both, or neither.
• Note every other business that appeared.
• Note what the assistant said you do, and whether it is right.
• Repeat each prompt at least three times in the same session.
• Repeat the whole set a week later before drawing any conclusion.
• Save the exact wording so next month’s run is comparable.
• Do the same on Perplexity and in Google’s AI Overviews.
• Only then consider whether a paid tool adds anything.

Is ranking in ChatGPT different from ranking in Google?

The underlying work overlaps more than the marketing suggests, but three things genuinely differ: there is no list of positions, rendering failures are far more punishing, and corroboration carries more weight relative to links.

Conventional SEO remains the foundation, not a legacy concern. The ChatGPT search surface draws on a web index, and a page that is not indexed cannot be retrieved from it. Every hour spent on indexation, site structure and technical health serves both channels at once. Anyone presenting AI visibility as a replacement for search work is describing a dependency as an alternative.

What differs is emphasis. Google renders JavaScript reasonably well; many AI crawlers do not, so a rendering failure that costs you a little in Google costs you everything here. Google shows ten results, so position two and position eight are both visible; an assistant shows three sources, so the tail does not exist. And Google’s ranking leans on links, while retrieval leans on whether the passage answers the question and whether anyone else corroborates it.

How does ChatGPT compare to Perplexity and Google AI Overviews?

Perplexity fetches live and cites densely, so technical fixes show up there first. Google’s AI Overviews lean on the Google index, so conventional SEO moves them most. ChatGPT sits between the two and is the largest audience of the three.

This matters for sequencing. If you want the fastest visible confirmation that your technical work landed, watch Perplexity, because it fetches at question time and cites almost everything it uses. If you want to move Google’s AI surfaces, the work is largely the conventional SEO you should be doing anyway, plus allowing Google-Extended. ChatGPT responds to both kinds of work, which is why the method on this page covers both.

It also matters for reporting. These are three different systems with three different behaviours, and a single AI visibility number that averages them together hides more than it shows. Measure them separately even when the work is shared.

Which of your pages are most likely to get cited?

Service pages, comparison pages and documentation. Specific, bounded, checkable content earns citations; broad thought-leadership almost never does.

The pattern across the sites we have audited is consistent. Pages that answer a narrow question with specifics get quoted. Pages that survey a topic at a high level get passed over in favour of a publisher who surveyed it better. This runs directly against the instinct to write big ambitious pieces, and it is worth resisting that instinct deliberately.

There is a commercial reading of this. The pages most likely to be cited are also the pages closest to a purchase decision, which means the citations you earn are disproportionately valuable. That is the opposite of the usual content-marketing shape, where the widest-reaching pages are the furthest from revenue.

Service pages

The single most valuable citation to earn, because the question behind it is commercial. Requires the page to actually describe the service in specifics rather than in adjectives.

Comparison pages

Cited often and scrutinised hard. Only worth building if you are willing to say something true about the alternative, including where it is better.

Documentation and technical explainers

The most reliably quotable content there is, because it is specific by nature. Underused by service businesses, who tend to keep this material in sales decks.

Pricing pages

Cited when the numbers are actually on the page. A pricing page that says contact us is not a pricing page as far as retrieval is concerned.

Case studies

Cited as evidence rather than as claims, which is why named and dated ones outperform anonymised ones by a wide margin.

General blog posts

Rarely cited, frequently produced. This is the category most content budgets are spent on and the one that returns least here.

How long does any of this take to show up?

Crawler access changes appear in logs within days. Rendering and structure changes take weeks. Citation frequency against a fixed prompt set moves over months. Anyone promising overnight results is describing something they cannot control.

The timescales are uneven because the mechanisms are different. Unblocking a crawler is an immediate change to a rule, and the evidence arrives with the next crawl. Re-rendering a template changes what bots read as soon as they refetch, which takes days to a couple of weeks depending on how often your site is crawled. Restructured content has to be refetched, re-evaluated and then selected, which is slower and not fully under your control.

The last stage is the slowest and the noisiest, and it is the one clients ask about first. Citation frequency is a rate, it varies run to run, and a real improvement only becomes visible once you have enough repeated measurements to see past the noise. Three months is a reasonable point to expect a readable signal on a site where the technical work was genuinely broken and is now genuinely fixed.

What does the engagement actually include?

Six deliverables: a crawler access report from your logs, a render audit per template, answer-first rewrites of your commercial pages, a source-attribution pass, a single entity description applied everywhere, and a prompt set with a recorded baseline.

Every one of these is a document or a change you can inspect. There is no proprietary score and nothing that only works while we are engaged. The prompt set in particular is yours from the first day, because a measurement instrument that you cannot run yourself is not a measurement instrument, it is a dependency.

The scope is driven by site size and template count rather than by a per-page rate, because the work is concentrated in templates. A two-hundred-page site with four templates is less work than a forty-page site with fifteen.

What should you ask a provider before hiring them for this?

Ask whether they want your server logs. It is the one question that separates people doing the work from people describing it, because the logs are the only artefact that settles anything.

Four follow-up questions do most of the remaining filtering, and they are worth asking in writing so the answers can be compared between providers.

Will you render our templates without JavaScript and show us the output?

This finds the expensive failure. A provider who does not run this test is guessing about the most common reason a site is invisible, and running it costs them ten minutes per template.

Which crawler user agents will you test, by name?

A vague answer about AI bots means nobody has opened a log file. The names are public and the list is short.

What will you measure before starting, and how often afterwards?

Without a recorded baseline taken against fixed wording, no later claim of improvement can be checked. The prompt set should be handed to you, not held.

What will you not do, and what will you not claim?

A provider who claims to influence every AI surface is overselling. The honest answer names the limits: no guaranteed placement, no paid shortcut, and different mechanisms on different assistants.

Who does the implementation?

Much of this work is developer work. A provider who produces recommendations and hands them to your team should say so up front, because that changes your cost and your timeline.

A provider who produces recommendations and hands them to your team should say so up front, because that changes your cost and your timeline.

Does this replace your SEO programme?

No, and a provider who says it does is selling you a dependency as a replacement. The ChatGPT search surface draws on a web index; a page that is not indexed cannot be retrieved from it.

The honest framing is that this is an extension of technical and editorial SEO into a channel with different failure modes. The overlap is large: indexation, site health, clear structure and genuine authority serve both. The additions are crawler access for a second set of bots, rendering without JavaScript, passage-level structure and a different measurement method.

Treating it as a separate discipline with a separate budget and a separate agency is how businesses end up paying twice for the same work and receiving two reports that contradict each other. It belongs inside the search programme, run by people who can read a log file.

What carries over unchanged

Indexation, crawl budget, site architecture, internal linking, page speed, and the general principle that genuinely useful pages outperform produced ones.

What is genuinely new

A second set of crawlers with their own rules, rendering requirements that are stricter than Google’s, passage-level structure as a first-class concern, and probabilistic measurement.

What gets more important

Corroboration from third parties, entity consistency, and the willingness to attribute claims rather than assert them.

What gets less important

Publishing cadence, keyword variants of the same page, and anything whose justification was that it fills a content calendar.

Publishing cadence, keyword variants of the same page, and anything whose justification was that it fills a content calendar.

What does this cost, and how is it scoped?

It is scoped like a technical audit: driven by template count and site size, not by a per-page rate. A fixed-scope audit with the six deliverables runs in one to two weeks once we have logs and staging access.

The reason scope follows templates rather than pages is that the failures cluster there. Testing one service page tells you what every service page does. The expensive variable is how many genuinely different templates a site has, how many of them assemble content client-side, and whether the fix is a configuration change or a build change.

Implementation is the other variable. Some clients want the audit and will action it with their own developers; others want the rewrites and the template changes done. Those are different engagements and should be quoted separately rather than bundled into a monthly retainer that obscures which one you are buying.

What makes it cheaper

Few templates, server-side rendering already in place, logs readily available, and a team that can action recommendations without a procurement cycle.

What makes it more expensive

A client-side-rendered site where the fix is a build change, a CDN managed by a third party, logs that have to be requested from a host, and a page inventory nobody has mapped.

What we would not charge for

Adding llms.txt, checking robots.txt, or running the ten-minute citation check described above. Those are on this page precisely so you can do them yourself before spending anything.

Can you do this in-house?

The technical half, yes, if you have developer access and can read server logs. The measurement half is where in-house programmes usually stall, because building and maintaining a fixed prompt set is dull, ongoing and easy to deprioritise.

Everything in the first five steps is within reach of a competent in-house team with developer support. The robots and WAF check is an afternoon. The render test is an afternoon. The answer-first rewrite is a real editorial project but a bounded one. The entity reconciliation is administrative.

What breaks down is the sixth step. A measurement set only produces value when the identical prompts are run on the same cadence for long enough to see past the noise, and that discipline is the first thing to go when a quarter gets busy. If you run it in-house, assign it to one person with a calendar reminder, and treat a missed month as a real cost rather than a gap you can backfill.

What an in-house team can do immediately

Check robots.txt and the WAF, fetch key templates without JavaScript, record a baseline against twenty prompts, and fix whatever the first two checks surface.

Where outside help usually earns its fee

Rendering changes on a framework nobody on the team owns, the editorial restructure across dozens of pages, and holding the measurement cadence steady for a year.

How to keep a provider honest

Ask for the logs and the raw prompt responses, not only the summary. If a report cannot be traced back to an artefact you could have produced yourself, it is a narrative.

If a report cannot be traced back to an artefact you could have produced yourself, it is a narrative.

What if you are a small business with a twelve-page site?

You are often better placed than a large one. The technical surface is small enough to get completely right, and specificity beats volume in this channel — which is the one competitive dynamic that favours the smaller site.

A large site has dozens of templates, a CDN somebody else configured, a content archive full of pages nobody will admit to owning, and a procurement process between a recommendation and a change. A twelve-page site has none of that. Every page can be rendered correctly, structured answer-first and attributed properly inside a fortnight.

The constraint on a small site is corroboration, not technique. If nobody outside your business has ever written about you, the fifth step is where your effort should go, and it is worth more than another ten pages would be. One named client case study hosted on the client’s own site is worth more here than a quarter of blog posts.

The small-site advantage

Complete technical correctness is achievable, which it rarely is at scale. Retrieval does not reward size; it rewards whether this specific passage answers this specific question.

The small-site constraint

Corroboration takes time and cannot be compressed. Start it before you need it.

Where not to spend

Bulk content production, AI visibility subscriptions, and anything sold as a shortcut. On a twelve-page site the entire technical pass costs less than a month of most of those tools.

Related reading on this site

Where each part of the method above is covered in more depth, and what to read next depending on which gate you are stuck at

• Answer engine optimization: the full service and method
• LLM SEO: the technical audit behind all of this
• AEO vs GEO vs LLM SEO: what the three terms actually mean
• Generative engine optimization
• GEO vs SEO: where they overlap and where they do not
• AI visibility best practices
• How AI is changing search
• Organic SEO services
• SEO audits and technical reviews
• Marketing strategy
• How long SEO takes to work
• How to rank higher on Google
• Local SEO services
• Local SEO vs national SEO
• Ecommerce SEO
• SEO content writing
• Content and commerce
• Web design and development
• Digital marketing services
• Google Business Profile optimization
• Organic search vs paid search
• Does URL length affect SEO?
• SEO myths worth retiring
• Affordable SEO services
• SEO for small businesses
• Digital marketing agency in New York
• SEO company in New Jersey
• PPC agency
• B2B social media
• Email marketing
• Branding agency
• Checkout optimisation
• Website redesign
• BigCommerce vs Shopify
• WordPress development
• About Progression Agency
• Client testimonials
• Talk to us about an audit

Where each part of the method above is covered in more depth, and what to read next depending on which gate you are stuck at

Frequently asked questions