Search Architecture

How to Use NovaX to Optimize for ChatGPT: A Universal Page Template

How to Use NovaX to Optimize for ChatGPT: A Universal Page Template

In This Article

    Knowing that OpenAI publishes crawler controls and no ranking formula is half the job. The other half is knowing what to fix on one specific page. This article is the page-level workflow: a seven-block template you can run against any URL, and how NovaX shows you which block is failing.

    In How to Optimize Your Website for ChatGPT I covered what OpenAI documents: which crawler does what, what the study data shows, and what is guesswork. I will not repeat that here. This article asks a narrower question. Given a single page, what do you look at, in what order, and how do you know afterwards whether anything changed?

    I use NovaX for that. It scores each page against engine-specific signals, one engine at a time. For ChatGPT, it models one particular idea: that the engine behaves like a semantic context and depth filter, one that rewards pages complete enough to answer the main question and the questions that follow it. That model is useful, and it is also a model, not a view inside OpenAI. I keep the two apart throughout, and the section on evidence levels says exactly where the line sits.

    What is in This Series

    This article is part of the cluster How to Measure Your Presence in AI Search Engines. The hub also contains AI Search Visibility Share of Voice, and a pair of articles for each engine:

    1. 1. The official guide articles say what the engine’s owner documents.
    2. 2. The measure articles say how to track presence over time.
    3. 3. The How to optimize with NovaX sits between them: it is the hands-on step where you inspect a page and decide what to change.

    What the ChatGPT Engine Is?

    ChatGPT is the semantic context and depth engine: it asks whether a page is substantial, whether it is organised, and whether it is unambiguous about what it covers. Those are the three questions the engine profile is built around, and every signal in the report maps to one of them.

    • Is this substantial? Is the page long and rich enough, with lists, tables and real detail, to answer the main question and the follow-ups a reader would ask next? This carries the largest weight, 40%.
    • Is this organised? Does a clean heading hierarchy show how the topics relate, so a section can be lifted out and used as an answer? This carries 30%.
    • Is this unambiguous? Do the named entities and the surrounding vocabulary make the topic clear, so a retrieval system can tell what the page is about? This carries the other 30%.

    None of the three is something OpenAI documents. OpenAI publishes crawler controls and, as far as I can find, no ranking criteria. The closest outside evidence is correlational: in the SE Ranking data I covered in the ChatGPT optimize article, longer pages and recently updated pages averaged more citations. That is a signal, not proof of cause. So the three questions are NovaX’s way of turning “comprehensive topical authority” into things a crawler can check on a page.

    Two points of scope matter before you read any score. First, “ChatGPT” covers search answers, pages fetched live for a user, and answers that never touch the web. The NovaX profile scores page-level signals that are relevant to the search side. It does not tell you which answer cited you, and it cannot see your firewall rules or your Bing Webmaster Tools account. The access checks in Step 1 of the optimize article still have to be done by a person.

    Second, the profile rests on a premise: that ChatGPT leans on Bing’s index. OpenAI says ChatGPT search uses third-party search providers, and its help page says the product sometimes partners with other providers and sends them rewritten queries. A vendor study from September 2026 (Peec) argues the real picture is wider, with OpenAI’s own index and several providers besides Bing. That is a vendor source, so weigh it accordingly. I treat the Bing premise as a reason to check Bing indexing, which is cheap, and not as a full account of where ChatGPT gets its pages.

    How to Read the NovaX ChatGPT Report

    Read two numbers first, then the signals underneath them, and never act on the headline alone. A ChatGPT report opens with an Engine Score out of 100 and a Citation Probability as a percentage. They answer different questions, and they can disagree.

    • Engine Score: how completely the page’s signals match what the ChatGPT profile looks for.
    • Citation Probability: a more cautious eligibility estimate. In the report I am using as the example, a page scored 98 out of 100 and 60% citation probability.

    Those two are not the same. A page can be well formed, deep and clearly organised, and still be a modest bet to be cited, because the things that separate cited pages from skipped ones are not all on the page. A high score beside a middling probability is not a bug to explain away. It tells you the page is in good shape and the remaining gap needs a different kind of work. When the two diverge, open the detail before you celebrate either number.

    The engine weights its score in three parts:

    ComponentWeightWhat it covers for ChatGPT
    Extractability / depth40%Word count, list items and tables. Substantive content is the primary signal in this profile
    Structural integrity30%Heading hierarchy. Multiple H1s, or no headings, indicate low authority
    Entity clarity30%Named entities, topical coherence and NLP-friendly prose

    The report also scores two further core signals, schema and metadata and freshness and decay, that appear beside these three without a listed weight in the breakdown. I read them as supporting context, and not as things that drive the score. “Primary ranking signal” is NovaX’s wording for the depth component, and OpenAI publishes no ranking formula, so I read it as NovaX’s modelling choice.

    The tabs then split the page into parts you can act on:

    TabWhat it tells youWhat you do with it
    OverviewScore, citation probability, top actions, issues, engine-specific checksDecide where to start
    SignalsThe five core signals scored individuallyFind the weakest block of the template
    SimulatorProjected score impact of each fixOrder fixes by payoff before you touch the page
    QueriesThe questions found on the page and how eligible each isSpot weak or noisy question headings
    EntitiesEntities detected, their type, schema status and clarityDecide which entities deserve attention
    RecommendationsSpecific issues raised by the InspectorFix the cheap, concrete ones
    FreshnessFreshness score and sensitivity for the engineDecide whether the page is due an honest update

    The Simulator is the part I use most. You tick the fixes you plan to make and it projects the score change, so you can see in advance whether a fix is worth an hour. It also has a state that is easy to misread: when every key signal is already in place, it has nothing to offer and says so. In the example report, the projection sat at 98 with no fixes to select. That is a result, and the worked example below returns to it.

    The Universal ChatGPT Page Template

    Seven blocks cover everything the ChatGPT profile checks, and they apply to any page on any site. Work through them in order. The first is about not locking yourself out, and the rest follow the engine’s weights: depth, then structure, then entities, then the supporting signals.

    1. Access: do not exclude yourself

    • Goal: OpenAI’s search crawler can reach the page, the page is indexable on the Bing side, and GPTBot is a decision, not an accident.
    • NovaX checks: whether GPTBot is blocked. The report header shows “GPTBot allowed”, and NovaX describes the rule as: block it and the page is removed entirely.
    • What OpenAI documents: its crawler documentation separates three agents. OAI-SearchBot is for search, and a site opted out of it will not be shown in ChatGPT search answers, though it can still appear as a link. GPTBot is about training. ChatGPT-User covers actions a user starts, where robots.txt rules may not apply. The settings are independent. So I read the NovaX GPTBot check as a sign that someone has or has not blocked OpenAI’s crawlers in general, and I confirm OAI-SearchBot separately.
    • You check by hand: the robots.txt rules for OAI-SearchBot and GPTBot, firewall and CDN rules that might block either, and indexing status in Bing Webmaster Tools. The Copilot measure article shows what that tool does and does not tell you. OpenAI says a robots.txt change can take about 24 hours to be reflected in search. For wider crawl and indexation problems, start with the Indexation and Crawl Diagnostic.
    • Done when: you can name the rule that governs each OpenAI agent on this site, and the GPTBot decision was made by the person who owns it. Allowing training is an intellectual property call, not an SEO call, as I explain in Everything You Need to Know About AI Intellectual Property.

    2. Depth: enough substance to answer and extract

    • Goal: the page holds enough real content to answer the main question and the obvious next ones, in forms that are easy to lift out.
    • NovaX checks: word count (300+ passes, and the report calls 600+ the sweet spot for citation eligibility), list items and tables. This is the 40% component, scored as data extractability in the Signals tab.
    • Fix: add substance, not length. A comparison table, a numbered procedure, a definition that was missing, a worked example. If a section can be deleted without anyone noticing, delete it before you add anything.
    • A rule I apply: do not treat 600 words as a target. In the SE Ranking data, longer pages did average more citations, but long and padded is still padded. The threshold is NovaX’s, not OpenAI’s.
    • Done when: every section does a job, and a reader would not need to leave the page to understand the answer.

    3. Structure: one H1, a clean hierarchy

    • Goal: the headings show how the topics relate, so a section can be extracted and read on its own.
    • NovaX checks: that the page follows H1, then H2, then H3 in a logical order. Multiple H1s, or no headings at all, count against the page.
    • Fix: one H1, with H2s for the main sections and H3s only inside them, with no skipped levels. Make each heading say what the section answers, and put the direct answer in the first paragraph under it. Check the rendered HTML as well as the editor. Themes and page builders sometimes add a second H1 through a logo, a hero block or a widget, and the editor will not show it.
    • Done when: the outline, read on its own, describes the page.

    4. Entities: surround the topic with what belongs to it

    • Goal: the page names the entities that naturally sit around its main subject, so the topic is unmistakable.
    • NovaX checks: entity co-occurrence (a person, an organisation and a product mentioned together) and an entity clarity score. The Entities tab lists what it detected, with type and clarity.
    • The profile’s own example: for a page about an electric truck, it looks for battery chemistry, charging infrastructure and fleet logistics. Their absence reads as thin coverage.
    • Fix: map the five to eight entities that naturally co-occur with your primary topic. Make sure each appears at least once in an informative sentence, not in a keyword list. For this very page, that set would be OAI-SearchBot, GPTBot, Bing, robots.txt and schema, among others.
    • A rule I apply: do not add an entity because it is on a list. If the page has no reason to discuss it, a forced mention lowers quality for the reader, and the profile is not worth that.
    • Done when: each entity is there for a reason a reader would see, and each is defined plainly the first time it appears.

    5. Breadth: answer the follow-ups on the same URL

    • Goal: a reader does not have to click away to ask the next obvious question.
    • NovaX checks: topical breadth, meaning whether the content covers likely follow-up questions within one URL, and the question-style headings it finds on the page in the Queries tab.
    • Fix: add a visible FAQ or “Related Questions” section for the three to five most likely follow-ups (cost, risk, alternatives, who it is for, how long it takes). The report says this single change can lift a ChatGPT score more than any structural fix. That is NovaX’s modelling claim.
    • A rule I apply: write the FAQ for readers, and do not treat FAQ markup as a ChatGPT tactic. In the SE Ranking study, FAQ schema was associated with fewer citations (3.6 against 4.2), and the authors suggested that reflects the kinds of pages that use it. A visible FAQ and FAQ schema are two different things.
    • Then read the list for noise. Navigation labels and footer links can be picked up as questions, and they should not be there.
    • Done when: every extracted question is one a real reader would ask, and each has a clear answer on the page.

    6. Schema and freshness: a clean baseline, honestly dated

    • Goal: the page carries a standard, valid structural baseline, and a real update history.
    • NovaX checks: that standard schema is present as a fallback (Article, Product or FAQ), with a valid author and datePublished as the minimum, plus the freshness score. The profile’s reasoning is that Bing leans on schema when parsing is ambiguous. I treat that as NovaX’s premise, not as documentation.
    • Fix: Article schema with a real author and publisher, datePublished and dateModified, all matching what is visible on the page. The skeleton is below. For the wider question of which markup is worth the effort, see Schema Markup for AI Search: What Actually Matters?
    • Rule: change the date only when the content changes. A date that moves without an edit teaches engines to distrust it, and in the SE Ranking data it was real recency, pages updated within three months, that correlated with more citations.
    • Done when: the schema date, the visible date and the actual last edit agree.

    7. Connections: internal links and image alt text

    • Goal: the page is tied into its cluster, and its images carry meaning in text.
    • NovaX checks: at least three internal links, and image alt text coverage. Links signal topical authority within the site, and alt text adds semantic density for text parsing.
    • Fix: link from the body, in context, with anchor text that says what the destination covers, and link to pages in the same cluster that a reader would actually want next. Write alt text that describes what the image shows and why it is on the page. Skip decorative filler.
    • Done when: every link has a reason, and no content image is missing alt text. Check the Recommendations tab as well, because a page can pass the coverage check and still list individual images to fix.

    The schema skeleton for block 6

    Adapt this, and do not paste it as is. Every value must match what is visible on the page.

    {
      "@context": "https://schema.org",
      "@type": "Article",
      "mainEntityOfPage": "https://example.com/your-page/",
      "headline": "Your page title",
      "datePublished": "2026-01-15",
      "dateModified": "2026-11-21",
      "author": {
        "@type": "Person",
        "name": "Author Name",
        "jobTitle": "Their real job title",
        "url": "https://example.com/author/author-name/"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Your Organization",
        "url": "https://example.com/"
      },
      "about": [
        {
          "@type": "Thing",
          "name": "A topic the page genuinely covers"
        }
      ]
    }
    

    The ChatGPT profile asks for the minimum: an Article with a valid author and datePublished. Everything beyond that is optional for this engine, and dateModified is there to keep the freshness signal honest.

    Ready to see how ChatGPT sees you?

    If you would like the access, depth, entity and content gaps scoped across a whole site, that is the work I do in Enterprise Search Advisory, and you can talk to me here.

    A Worked Example: One Page, One Report

    A page can score 98 out of 100, pass every check and leave the Simulator empty, and the skill is in knowing what that does and does not tell you. Here is a real NovaX ChatGPT report for one page, my own ChatGPT optimization article. The specific numbers will differ on your pages. The reading order will not.

    The headline: Engine Score 98, Citation Probability 60%, GPTBot allowed, and all six engine-specific checks passing. The five core signals read structural integrity 100, data extractability 95, entity clarity 100, schema and metadata 100, and freshness and decay 88. All three research-backed signals (entity co-occurrence, topical breadth and the schema fallback) show as present. The score and the probability disagree, so I read the detail.

    What the report showsWhat I make of itActionPriority
    Last Inspector run was 4 days agoThe numbers describe a recent version of the pageNo re-inspection needed unless the page has since changed1
    Simulator: “All key signals are already in place”, projection stays at 98No page-level fix is projected to move the scoreDo not edit the page to chase the last two points2
    Freshness 88 and data extractability 95, the two lowest core signalsNeither is broken; both have a little headroomImprove them when the page is next materially edited3
    Entity co-occurrence, topical breadth and schema fallback all presentBlocks 4, 5 and 6 are satisfiedProtect them: do not lose them in a redesign or a template change4
    Score 98 against citation probability 60%The eligibility estimate sits well below the scoreLook beyond the page: authority and third-party mentions5
    Queries and Entities tabsA passing checklist does not guarantee the extracted questions and entities are cleanOpen both and read the lists6

    Three things in that table are worth saying out loud.

    An empty Simulator is a result, not an error. It means the profile has no more page-level fixes to project. It does not mean the page is finished, and it does not mean the work is done. It means the remaining gap is not one this tool can show you on the page.

    A page-level 98 is not a citation forecast. The 60% beside it is the honest half of the report. In the study data I covered in the optimize article, the strongest correlates of ChatGPT citations were off the page: referring domains, domain trust and third-party mentions. That is my inference from correlation data, not something the report states. But it fits: a page that is deep, organized and unambiguous has done what the page can do. What is left is slow authority work.

    Do not chase 100. The last two points are not worth an edit that makes the page worse for readers. If the freshness number is the lowest, the right fix is a real update when the content changes, not a date that moves.

    The result of working a list like this is a page that is protected from regressions and a clear sense of where the next effort belongs. It is not a guarantee of citations, which is the subject of the next section.

    What NovaX Measures vs What OpenAI Documents

    A NovaX score is a model of what tends to go with eligibility. It is not OpenAI’s own weighting, and nobody outside OpenAI has that. I built NovaX, so you should read my claims for it with that in mind. No third-party tool has access to how ChatGPT selects sources, and that applies to NovaX as much as to any other tracker.

    So the weights in the report, such as the 40% on depth, are NovaX’s modelling choices. Here is how each part of the template stands against what OpenAI says and what the outside data shows, as I read it from the sources I cited in the optimize article.

    Template blockOpenAI’s position or the outside evidenceNovaX’s treatmentMy read
    1. OAI-SearchBot allowedDocumented: it controls appearance in ChatGPT search answersNot checked directly; the check is on GPTBotDo first. The strongest evidence of any block. Verify by hand
    1. GPTBot decisionDocumented: relates to training, and is independent of searchPass or fail engine checkA legal call as much as an SEO call
    1. Bing indexingOpenAI says search uses third-party providers, Bing among them. One vendor study reports a wider mixThe profile’s premiseA cheap check, but not the whole story
    2. Depth (length, lists, tables)Not documented. Correlation in the SE Ranking study (under 800 words averaged 3.2 citations, over 2,900 averaged 5.1)Weighted 40%Do it honestly. Add substance, not words
    3. Heading structureNot documented. Vendor practiceWeighted 30%Cheap. Do it
    4. Entity co-occurrence and clarityNot documentedWeighted 30%Plausible, and cheap if the mentions are natural
    5. Follow-up coverageNot documented. FAQ schema was associated with fewer citations in the studyCalled the biggest single liftWrite a visible FAQ for readers. Skip FAQ schema as a tactic
    6. Schema baseline and freshnessNo OpenAI statement. Updated within three months averaged 6 citations against 3.6Standard schema is the fallback signalCheap and honest if the dates are real
    7. Internal links and alt textNot documented3+ links, alt text coverageStandard site and accessibility practice

    What a score can and cannot do follows from that table. It can show you whether a page is complete against a defined checklist, order your fixes by projected payoff, catch regressions after a template change, and compare pages on the same site. It cannot prove that a fix caused a citation, show which answer used the page, or see your robots.txt, your firewall or your Bing account.

    The Fix-and-Verify Loop

    Run the loop on one page first, then on the pages that matter most, and keep each round small enough to attribute. Changing seven things at once tells you nothing about which one worked.

    1. Inspect. Run the page through NovaX and open the ChatGPT report. Note the Engine Score, the Citation Probability and the date of the inspection.
    2. Check access by hand. Confirm the robots.txt rules for OAI-SearchBot and GPTBot, look at firewall and CDN rules, and check indexing in Bing Webmaster Tools, because the report cannot see any of them.
    3. Simulate. Use the Simulator to see which fixes carry the highest projected impact, and tick the ones you plan to make. If it is empty, skip to step 7.
    4. Fix in template order. Work through blocks 1 to 7, cheapest and most concrete first. Fix site-wide problems, such as a second H1 from the theme or navigation being read as questions, in the template and not page by page.
    5. Request re-indexing for pages with major changes. Given the Bing premise, Bing Webmaster Tools is the place. Allow time: OpenAI’s own note is about 24 hours for a robots.txt change to take effect in search.
    6. Re-inspect. Run the page again and compare. Did the blocks you touched improve, and did anything else regress?
    7. Measure over time. A score changing is not the same as being cited more, and ChatGPT can return different sources for the same prompt, so one check proves little. For that, track presence over repeated runs with the method in How to Measure Your Presence in ChatGPT, and see AI Search Visibility Share of Voice for the wider picture.

    Two supporting reads help at the this stage: How to Write a Blog Post for AI SEO and Citations and AI Content Structure for Enterprise Visibility. For the schema side in more depth, see Schema Markup for AI Search: What Actually Matters?

    Where to Go From Here?

    Start with the AI Visibility Inspector on one page you care about, and work the seven blocks in order. When you need ChatGPT tracked across your pages and against named competitors over time, NovaX covers it. And if you want the access, depth, entity and authority gaps scoped for your organization, contact me.

    What This Is NOT

    This is not a promise that a higher score earns more citations. It is not a replacement for the access checks, which no page scanner can do for you. It is not a claim that OpenAI uses NovaX’s weights, because it does not publish its own. And it is not a case for polishing a page when the real gap is authority, since the strongest correlates in the data I have seen sit off the page.

    It is a way to inspect one page against a defined checklist, fix the cheap and concrete gaps first, and confirm afterwards that the page actually changed.

    FAQ

    It is the engine profile that scores a page on semantic context and depth: whether the page is substantial, organised and unambiguous about its topic. It scores page-level signals. It does not tell you which ChatGPT answer cited you.

    The Engine Score shows how completely a page’s signals match the ChatGPT profile. The Citation Probability is a more cautious eligibility estimate. They can differ, as a 98 score beside a 60% probability shows, and the gap is a reason to read the detail and to look beyond the page.

    NovaX checks GPTBot, but OpenAI documents it differently. Per OpenAI, OAI-SearchBot is the agent for ChatGPT search, and GPTBot relates to training. The two settings are independent. Check OAI-SearchBot yourself, and treat the GPTBot decision as an intellectual property call.

    OpenAI says its search uses third-party search providers, and Bing is one it has named. A vendor study from September 2026 argues the mix is wider and includes OpenAI’s own index. Checking Bing indexing is cheap, so do it, but do not assume it is the only route.

    NovaX treats follow-up question coverage as a major signal, and a visible FAQ is good for readers. FAQ schema is a separate matter: in the SE Ranking study it was associated with fewer citations, which the authors suggested reflects the kinds of pages that use it. Write the FAQ, and do not sell the markup as a tactic.

    No. The report passes a page at 300+ words and calls 600+ the sweet spot, but those thresholds are NovaX’s, not OpenAI’s. Length correlated with citations in the study data. Padding does not become useful because the count is higher.

    Because every key signal is already in place, so it has no fix to project. That is a result, not an error. It means the remaining gap is not something a page-level edit will close.

    After any material change, and before acting on a report that is old. A report from four days ago describes a recent page. One from months ago does not.

    No. You can promise a baseline, a fixed set of changes and repeated measurement.

    This article was researched and drafted with the assistance of AI tools and reviewed and edited by author prior to publication.

    Share in 𝕏 ✉
    Ivica Srncevic
    Author

    Ivica Srncevic is an independent AI strategist, researcher, framework author, and international speaker focused on AI sovereignty, knowledge infrastructure, governance, AI retrieval, and the evolving relationship between organizations and intelligent systems. His work examines what AI systems can see, retrieve, infer, and reconstruct from organizational information, and how organizations can retain greater control over their data, knowledge, and AI infrastructure. In 2026, he spoke at the AIFOD Geneva Summit at UN Geneva on what nations must own and what they can safely share, with a particular focus on data ownership, control, and sovereign AI infrastructure.

    Articles: 275