IT Master

What to demand from a generative engine optimization provider

Some GEO retainers sell a promise nobody can keep: guaranteed citations in ChatGPT or Perplexity. Here is what a serious provider measures, ships and reports, so you can tell the two apart before you sign.

Published 2026-09-02, reviewed 2026-09-04, by Rodd Azad, editor-in-chief. How we check facts.

What you are actually buying

Generative engine optimization services exist to get a site cited when someone asks ChatGPT, Perplexity, Gemini or Google's AI Overviews a question in your market. The assistant retrieves a handful of pages, lifts the passages that answer the question, and names its sources. GEO is the work of becoming one of those sources.

That work is less mysterious than the sales decks suggest. Assistants reach pages through ordinary web search and their own crawlers, so a page blocked to those crawlers or missing from the index cannot be cited at all. A slow page is a weaker candidate, not an impossible one. In practice the passages that get lifted answer a question directly, carry specifics a model can attribute, and sit on a site that covers the topic repeatedly rather than once. Most of what moves the needle is conventional publishing done to a higher standard, plus measurement that ordinary SEO tooling does not provide.

The parent guide, generative engine optimization, covers the mechanics. This page is narrower: how to judge a provider offering to do it for you.

In one sentence

Generative engine optimization services are the work of getting a site retrieved and cited by AI assistants such as ChatGPT, Perplexity and Google AI Overviews, measured against a fixed set of buyer queries.

What a good provider measures

Ask any vendor one question first: measured against what? A credible answer names a fixed set of queries your buyers actually ask, checked on a schedule across named assistants. Without that baseline, every later claim is unfalsifiable.

The minimum measurement set:

  • Citation rate — for each query in the set, whether your domain appears among the cited sources in GPT and Perplexity, recorded on a fixed schedule, with the misses kept.
  • Which page was cited — so you learn which formats earn retrieval and can repeat them.
  • AI crawler access — whether OAI-SearchBot, ChatGPT-User, PerplexityBot and ClaudeBot can fetch your pages at all. Google-Extended belongs in a different column: it is not a fetcher, it governs whether pages Googlebot already has are used to train Gemini, and blocking it does not remove you from AI Overviews. Our free AI crawler check shows this.
  • Search Console — impressions, clicks and indexation for the same pages, because retrieval starts with being indexed.

Notice what is missing: rankings for vanity keywords and screenshots of one flattering prompt. A single conversation proves nothing; assistants vary their sources between sessions. Only a repeated query set, run the same way each time, turns citations into a number you can trend. Take your own baseline before any sales call with the AI visibility checker.

In one sentence

The minimum honest measure of a GEO service is a fixed set of buyer queries, checked on a schedule across named assistants, with the misses recorded as well as the hits.

What a first-month baseline must contain

Everything a GEO provider claims later is measured against the baseline they take now, so this is the part to be difficult about.

  • The query set, published before work starts. Fixed, written down, and yours to keep. If it changes between reports you are being shown the wins, not the trend.
  • Retrieval, not recall. The probe must run the assistant's web-search mode. We learned this expensively on our own estate: our probe asked a plain question with no search and returned 0 citations from 80 probes across four sites, for five consecutive weeks. With search switched on, the same questions on the same day named two of those sites first and second, with citation annotations to their own pages. Both numbers were honest; only one was about our pages.
  • No brand-name queries. Ask a model to list suppliers for a query that is your own brand name and it will name your brand. Our earlier figures looked healthy for exactly that reason. Curated queries produced an honest zero, which was progress.
  • Answers counted, not links. A publisher quoted three times in one answer is one endorsement.
  • Errors stored as errors. "We could not measure" and "nobody was cited" are different claims; a failed probe recorded as a zero manufactures a decline.
  • Per-engine reporting. On 31 August 2026 the same four sites, same twenty questions, same day, were cited by Perplexity in 12, 7, 5 and 2 answers out of 20, and by ChatGPT's web search in 10, 5, 3 and 0. One blended score hides that entirely.
  • The indexed share of your existing pages. Retrieval starts from a conventional index, so this bounds everything else.

Cadence is a cost decision and should be stated as one. A single search-backed ChatGPT probe costs roughly 8,800 input tokens plus a billed web search at about two cents, and takes about a minute. Nightly probing of twenty questions across four sites is roughly $54 a month to re-measure something that moves in weeks. We run ChatGPT weekly and Perplexity nightly, and skip rather than downgrade on the other days, because two different methods in one column produce a trend line that means nothing.

In one sentence

Ask one question before anything else: does your visibility probe use the assistant's live search, or does it ask the bare model? A probe without retrieval measures how famous a brand already is, and it will report a young site as invisible however well its pages are built.

What a good provider ships

Measurement is the frame; the deliverable is pages. A provider that reports beautifully but publishes little is selling you a dashboard.

Every piece that leaves a serious GEO shop should carry:

  • A direct answer near the top of each section, in one or two sentences an assistant can lift without editing.
  • Facts that exist nowhere else — your product data, datasheet figures, the support questions your team answers every week. Readers and assistants both prefer the source over the paraphrase, and a competitor cannot copy what they do not have.
  • Structured data — Article, FAQPage and, where relevant, Product JSON-LD — plus an FAQ block written for the question the query actually asks.
  • Internal links into the rest of the cluster, so one cited page lends authority to the pages around it.
  • Clean technical delivery — indexable, fast, a canonical URL, and a ping through IndexNow when it goes live.

Then, cadence. Publish steadily rather than dumping a quarter's pages in one week: a drip is what a crawl budget and an editorial calendar can both absorb, and it gives you a clean before-and-after to read. A refresh schedule matters too, because an answer that was right in March and wrong in September gets dropped as a source.

What a good provider reports

A useful GEO report fits on one page and answers four questions: what did we publish, what did it cost, what is being cited, and what changed since last month.

Look for:

  • The fixed query set, unchanged month to month, with citation counts per assistant. If the set changes every report, you are being shown the wins.
  • Per-page results tied to Search Console, so a citation can be traced to a URL and a URL to clicks.
  • Cost per published page, not per draft, so you can see what a citation actually cost to earn.
  • A list of what was rejected or rewritten and why. A provider that never rejects its own work is not checking it.
  • What will be refreshed next, based on decay in the numbers rather than a hunch.

Be wary of reports built around "share of voice" indices without the raw counts behind them, and of any metric the provider invented and only they can compute. You should be able to reproduce the headline number yourself by asking the assistant the same questions.

What you can audit yourself, without the provider

None of these needs access to a vendor's dashboard, and together they catch most of what goes wrong.

  • Read your robots.txt the way a crawler does. Fetch the file and find the group whose user-agent matches the bot you care about. A crawler obeys only its most specific matching group and ignores every other one, so rules added to the wildcard "User-Agent: *" group never reach a bot that has its own group. We shipped exactly that mistake: facet rules in the wildcard group, thirteen named crawler groups on each site, and over a 24-hour window 382 of Googlebot's 1,437 requests — 26.6% — went to faceted URLs we believed were blocked.
  • View the served HTML, not the rendered page. If the body copy only appears after JavaScript runs, some fetchers will not see it.
  • Run one URL through Google's Rich Results Test. Compare the questions it extracts with the questions on the page. Schema that disagrees with the copy is worse than no schema.
  • Check Search Console's page indexing report. What share of the pages published for you is indexed? "URL is unknown to Google" with no crawl date is a discovery problem, not a quality one, and the remedy is IndexNow, internal links and a truthful sitemap lastmod, not a rewrite.
  • Ask an assistant a question yourself and look for links. An answer with citations came from retrieval and is evidence about your pages. An answer with no links came from the model's memory and is evidence about your brand's fame.
  • Check your analytics for AI referrals. In our own analytics, the links ChatGPT cites arrive carrying utm_source=openai; Perplexity referrals arrive with perplexity.ai in the referrer. Both undercount, because a citation nobody clicks leaves no trace.

If a provider's report and these checks disagree, the checks are the ones you can repeat.

Red flags that should end the call

Some signals are disqualifying on their own:

  • Guaranteed citations, positions or traffic. Nobody controls what an assistant retrieves. A provider who promises it is either guessing or planning to move the goalposts.
  • Screenshots as proof. One prompt, one session, one lucky answer. Ask for the query set and the miss rate instead.
  • No mention of crawler access. If the pitch never checks whether GPTBot can fetch your site, the vendor has skipped the first gate.
  • The same model writes and checks. A model grading its own draft finds the errors it did not notice the first time, which is to say very few.
  • Charging for drafts you never see. You should pay for pages that go live, not for attempts.
  • Twelve-month lock-ins for work that is measured monthly. If the numbers are real, the provider does not need the contract to keep you.
  • No takedown control. Content published in your name should be yours to amend or remove, in your own CMS, without raising a ticket.
  • Claims that Google or OpenAI prefer AI-written content. Neither has said any such thing, and a provider who tells you otherwise is inventing an endorsement.

None of this is exotic; it surfaces in the first ten minutes of a call if you listen for it.

When GEO services are not worth buying yet

Some of the enquiries we get should not become customers this quarter, and saying so is cheaper for both sides than a disappointing engagement.

  • Your pages are not indexed. Every assistant in wide use retrieves from a conventional index. Across four sites we publish for, measured on 25 August 2026, indexed share ranged from 25 of 27 to 12 of 25 on one pipeline — so this is not a rare edge case. Under about 80%, spend the money on discovery.
  • There is no first-party data and no search demand. Grounded articles need something to be grounded in. Without a catalogue, datasheets, measurements or a support inbox, the output converges on the pages already ranking, which is precisely what Google's spam policy warns about when it describes many pages generated without adding value for users.
  • Your buyers do not ask assistants. Some markets are transactional or navigational and produce almost no cited answers. Sample twenty real buyer questions before committing; if none of them returns a cited answer at all, the surface is not there yet.
  • You need links, PR or a rebrand more than pages. Content compounds on top of authority; it does not substitute for it.
  • You cannot measure. Without Search Console, analytics and a probe you trust, an engagement has no feedback loop and neither side can tell whether it worked.
  • You need a guarantee. We do not give one, because nobody controls what a model says, and a supplier who does is describing an engine's behaviour as though it were their own.

In one sentence

If your existing pages are not being indexed, buying generative engine optimisation is buying the fourth stage of a four-stage chain. Fix discovery first; it is cheaper and it gates everything after it.

How IT Master approaches it

IT Master is an autonomous engine, not a retainer, and the checklist above is largely what it was built to satisfy.

Topics come from real demand: DataForSEO keyword data and your own Search Console. Research reads competitor pages, forums and your first-party data — product records, datasheets, support Q&A — so drafts carry facts a rival cannot reproduce. A Claude model writes. Then Gemini and GPT models judge it: fact-check, novelty against competitors, an E-E-A-T critique and AI-tell detection, with up to three revision rounds. The model that writes is never the model that checks.

The judges are strict on purpose. Across 186 Standard-tier runs, 21% of drafts cleared every check on the first attempt with no revision, and 51% got there within the three revisions allowed; the rest never published. Pro-tier judges are stricter: 25% cleared every check across a small sample of 16 runs.

Passing articles publish to your site with JSON-LD, internal links, a hero image and an FAQ; IndexNow pings search engines. A nightly loop refreshes decaying pages and measures Search Console alongside citations in GPT and Perplexity. You pay per published article from a prepaid balance, with no subscription; see pricing. Delivery is a WordPress plugin, a Next.js SDK, a REST pull feed or a signed webhook. It runs for live UK CCTV trade and retail sites.

Questions to bring to the sales call

Take these into any conversation about GEO services, including one with us:

  • Which queries will you measure, and can I see the list before we start?
  • Which assistants do you check, how often, and do you keep the misses?
  • Who checks the writer's work, and is it a different model or a different person?
  • What do I pay for a draft that fails your own checks?
  • Can I remove a published page myself, today, without raising a ticket?
  • Which of my own data will the content draw on that a competitor lacks?
  • Can you show me the JSON-LD and the crawler access on a page you published last month?

A good provider answers each in a sentence. A weak one changes the subject.

If you want a baseline before you talk to anyone, run the free site check on your domain, then try the AI visibility checker for the citation side. The numbers they give you are the ones to hold every vendor to, this one included. How it works sets out the full pipeline if you would rather see the mechanism than the pitch.

Sources

Primary sources for the claims on this page, all read on 4 September 2026.

Questions people ask

What do generative engine optimization services include?

A complete GEO service covers three things. Measurement: a fixed set of buyer queries checked on a schedule across named assistants, with citation counts per page. Publishing: pages written to be retrieved, with direct answers, first-party facts, FAQ and Article schema, internal links and clean indexation. Reporting: what shipped, what it cost per published page, what is cited and what will be refreshed next. Technical checks such as AI crawler access and IndexNow submission sit underneath all three. Anything that offers only one of the three is a tool or a dashboard, not a service.

Can a provider guarantee citations in ChatGPT or Perplexity?

No. Assistants choose sources per query and per session, and their retrieval changes without notice. A provider can raise the odds by making pages fetchable, answering questions directly and publishing facts other sites lack, and it can measure whether the odds improved. It cannot promise an outcome it does not control. Treat any guarantee of citations, positions or traffic as a reason to walk away, and ask instead for the query set, the miss rate and the month-on-month trend.

Should I hire a GEO agency or run a content engine?

It depends on where your constraint sits. An agency earns its fee on strategic, low-volume work: positioning, a handful of high-stakes pages, an awkward migration. An engine earns its keep when the constraint is throughput and consistency, because publishing every week for a year is what builds the depth of coverage assistants retrieve from. Nothing stops you doing both. Keep the agency for strategy, let an engine produce the recurring pages, and hold both to the same query set and the same Search Console numbers rather than to two different scorecards.

How do I verify a provider's citation claims?

Ask for the raw query list, then run a sample of it yourself. Put the same questions to ChatGPT and Perplexity, note whether your domain is named as a source, and compare your count with theirs. Expect variation between sessions, which is exactly why the list has to stay fixed and the misses have to be recorded. Then check the cited URLs in Search Console, because a citation pointing at a page with no impressions deserves a question. If a provider will not hand over the query list, there is nothing to verify.

What happens to the pages if I stop?

They stay where they are. Articles publish into your own CMS, on your own domain, so they do not vanish when the work stops. With IT Master there is also no subscription to cancel: you pay per published article from a prepaid balance, so stopping means not topping it up again. What does stop is the upkeep, meaning the nightly refresh of decaying pages and the citation measurement, and that is the part worth keeping running, because an answer that quietly goes stale also quietly stops being cited.

What does IT Master charge for GEO?

There is no GEO retainer and no subscription. You pay per published article from a prepaid balance, and a draft that fails the validation checks is not charged. There are four quality tiers: Lite, Standard, Pro and Premium. The higher tiers are judged against a stricter bar, which is why fewer of their drafts pass first time and more get sent back for revision. Measurement of Search Console and AI-assistant citations runs nightly as part of the pipeline. Current rates are on the pricing page.

How can I tell whether a GEO provider's visibility score is meaningful?

Ask whether the probe uses the assistant's live web search or asks the bare model, and whether the answers carried links. A probe without retrieval answers from the model's memory, so it measures how famous your brand already is rather than what your pages say. Ours did exactly that and returned zero citations across eighty probes for five weeks; with search enabled, the same questions on the same day returned real citations. Then ask whether the query set is fixed, whether brand-name queries are excluded, and whether failed probes are stored as errors rather than zeros.

Should I buy GEO services for a site that is not indexed yet?

No. Every assistant in wide use retrieves candidates from a conventional web index before citing anyone, so an unindexed page cannot be quoted whatever its structure. Check Search Console for the share of your pages that is indexed. Under roughly 80%, spend the budget on discovery instead: a truthful sitemap lastmod, internal links from pages that are crawled often, and IndexNow. We measured four sites on one pipeline in the same week and indexed share ranged from 25 of 27 down to 12 of 25, so this is common rather than exotic.

Do ChatGPT and Perplexity need separate work?

They need separate measurement, and usually the same pages. On 31 August 2026 we put the same twenty buyer questions to both engines for four sites on the same day: Perplexity cited them in 12, 7, 5 and 2 answers out of 20, ChatGPT's web search in 10, 5, 3 and 0. Perplexity cites many more sources per answer and is generally where a newer site appears first. The crawler permissions differ too, and are set independently in robots.txt, so a provider reporting one blended AI visibility number is hiding the more useful half of the picture.

Read next

See what search engines and AI assistants find on your site

Free, and no account. Type your address and the check tells you what is missing and what we would write first.

Check my siteSee what an article costs
Run the free site checkPricing