CitationOne now works inside AI assistants. Connect it to Claude Code, Claude Desktop or Cursor over MCP, with the same API key you use for integrations. In the conversation you can ask for an audit of a URL or a draft, see the score against the top 10 and the most important recommendations, browse earlier audits of a client and check your credit balance. An audit costs 1 credit, just like in the dashboard. You can also add the CitationOne skill to Claude: the assistant asks for approval before an audit, merges several pages of one client into one list of fixes and prepares a summary for the client. Setup instructions are in the API documentation, in the MCP section.
The Schema.org card in the audit report now has a "Show ready-to-paste JSON-LD" button. It opens a window with a <script> block built from your page data: headings, description, breadcrumbs, FAQ questions, logo and social profiles. The code is consistent with the markup you already have - it reuses its identifiers, and an incomplete block comes back completed, ready to swap in. Each block has a status: "ready" or "needs input", with the fields to fill in. You also get company and website code (Organization and WebSite) shared by all pages - when the site already has these blocks, the panel says so. Check the code with one click in the Rich Results Test. In Schema Gaps the ready code is available for every page and in the JSON-LD export. The button appears in new audits.
On October 30 Google changed how it provides page addresses in search results. For part of that day, audits fetched an incomplete set of top results: 1-3 pages were compared instead of 8-10, and some competitor addresses led to an error page. We fixed how results are read - the audit once again compares your content with the full set of top results, and competitor and AI Overview source addresses point to the right pages. If your audit from October 30 shows only a few competitors, run it again to get the full comparison.
The Competition block now has a Google AI Mode section. It shows whether AI Mode cites or mentions your page, which query we checked and which pages the answer draws on. The ChatGPT section now shows the answer as a chat user sees it, plus new information: the pages ChatGPT looked at but did not cite. When your page is among them, the status reads "viewed, not cited" - a sign that the model knows your content while the citation goes to a competitor. Both sections are also included in the PDF and Markdown exports.
Since mid-September Google has increasingly shown automated tools different results than people see: a page that looks correct but carries a swapped list of results. The audit then compared your content with competitors you do not see in search. We changed how search results are fetched - the top 10, "People also ask" questions, AI Overview, the map and products now come from a view that matches what a Google user sees. Competitors, topic gaps and recommendations therefore refer to the pages you actually compete with. Previously saved audits stay unchanged - to compare against the current leaders, run a new audit.
You can now arrange the recommendation list by article section. The "Content sections" switch groups recommendations in the order of the text: from metadata and the lead through each heading, and every card shows its priority. Sections of your article and individual recommendations have a "Preview changes" button - within seconds you see the section with the recommendations applied, with new and changed passages highlighted. A recommendation you hide is left out of the preview. Each audit includes 10 free previews. Apply everything as before in the "Apply to content" tab.
Apply to content now adds figures given in the audit recommendations to new sections, e.g. a duration or a typical cost, when the sentence covers general knowledge about the topic. Each such sentence appears on the "Figures to verify before publishing" list, so you can confirm the value against a current source. Values you stated on the page yourself - availability, price, catalogue number - stay in your wording. When a section's opening sentence needs correcting, a new one is written from the facts on your page. A change that turns out to have no effect after the final check is reverted and your original text returns. Long runs with many recommendations now finish in full.
When you enter a page address on citationone.com, we start working right away. Already on the sign-up screen you see the title of your page, its word and heading counts, the selected keyword with its search volume, and a list of what the recommendations will include. After you create the account, the audit starts on its own - no retyping the address and no extra click. When you arrive with an address while logged in, the wizard fetches the content and picks the keyword by itself, and you check the keyword and click the pulsing "Start audit" button.
Competitor content length in the report is now a median, with the range that holds the middle half of the top-10 pages shown below it. A single very long page no longer inflates the result, so the number better reflects how much the pages in the search results actually write. When there are too few pages to compare, or their lengths are extremely different, you see a dash with an explanation in place of a number. The expected level of detail is now calculated from the length of the top-10 pages, and the recommendation on angle and depth gives the typical page length in the results. The Effort Score check for content length uses the same median. The median appears in the report and the PDF, including saved audits, and in new bulk audit summaries.
Topics in Content Pruning now have clear names, e.g. "Buying guides", in place of keyword fragments. The split into topics is more stable: a new analysis after small changes on the site puts pages into the same topics, and their number follows the site's content. A cannibalization group now covers pages that are close to one another, so the list is shorter and more specific. Each topic shows its central page, the one that best represents it - a good candidate for a pillar page. The similarity threshold used in the analysis is shown above the groups - it rises on its own when the site's pages are built on one template. We also fixed reading of headings broken over several lines, where words used to run together. Earlier analyses display as before.
Content Pruning now shows the pages that stand out from the other pages of their topic. For each of them AI suggests an action: keep, expand, merge with another page, redirect or remove, and for merge and redirect it names a target address from your own pages. Side topics appear together in the summary, while borderline pages and pages between topics have their own lists. The CSV export includes the verdicts and target addresses. The analysis settings now hold only the cannibalization threshold. We also fixed meta description reading, which on some pages picked up text from a neighbouring tag. Earlier analyses display as before.
Automatic keyword detection now recognises the language of the content before proposing a phrase. Previously a German page ordered through the API without a declared language could receive a Polish keyword, and with it search results and search volume from the Polish market. We now read the language from the content itself - Polish, German and English - and your own "language" field takes precedence. The same fix covers bulk audits, which until now treated every German page as English. Audits ordered in the app were already correct and are unchanged.
The report has a new dimension: Page metadata. We check the title - its length, whether it carries the keyword and where, whether it names the central entity, whether it is written in capitals and whether the keyword repeats - and the meta description: whether it exists, fits the search result, carries the keyword and avoids characters that truncate the snippet. A brand name in the title is flagged only when it pushes the title past the character limit. The score feeds into CQS alongside the other dimensions, weighted higher for pages competing for the click: listings, product pages and landing pages. Next to the title and the description you can see how much CQS rises once the proposed version is live. For pasted content with no URL the dimension is not scored.
The wizard now has three steps instead of four. The screen for confirming the detected search intent is gone - you enter the address and the keyword, click "Start audit" and go straight to the work. We detect the intent, the context and the central entity on our side, and you see the result together with the alignment with the Google top ten on the progress bar, before the report is ready. The full comparison against the SERP consensus stays in the report. A keyword with no organic results is rejected before the credit is taken, so you never pay for an audit without a competitive comparison.
Some recommendations you put aside on purpose. Every recommendation card now carries a crossed-out eye icon - one click collapses it into a single greyed-out line, right where it sat in the list. You can see what you set aside, and one more click brings the full card back. A hidden recommendation drops out of the report you share by link, out of the exports and out of the public API, and the Apply to content tab stops listing it - so it will not reach the article even when you select everything. The counters next to each priority show active recommendations, and the CQS score stays as it was, because it describes the page itself.
The Report Builder in the Export tab now configures the public link. You set your company name and logo URL and tick what the recipient sees: the Summary, Dimensions, Recommendations and Apply to content tabs, and the cards inside them. The list covers only sections present in this audit. The settings apply to every link from your account, including ones already sent, and hidden sections leave the link together with their data. The header shows your logo and a "Report prepared by" line, or the CitationOne logo. PDF export and the HTML preview are off for now.
Information Gain compares your claims against the top ten results. Until now a claim that shared vocabulary with competitors was filed as "they already have this", including when it said the opposite. Numbers suffered most: after changing only its value, a sentence still passed as a repetition in 98 cases out of 100. Every borderline match is now checked for meaning, in both directions: a claim ruled unique can be demoted, and one ruled a repetition can be recovered. Passages copied word for word remain repetitions. The effect is visible: Information Gain in new audits can come out higher than in older audits of the same page, and we do not recalculate old results. The card also shows how many claims carry hard data (a number with a unit, a date, a trend) and how many of those are unique.
The audit report now has two views. Advanced is the panel you know and stays the default. Simple mode shows the path to getting work done: the page score at the top, the recommendations, apply to content and the export. The per-dimension analysis and the extra tools (Clustering, Pruning, Schema Gaps, Internal Links) stay in advanced mode. You switch modes in the sidebar footer and on the login screen, and the choice is stored on your account. The TOP 10 ranking sits below the recommendations there. Exports contain the full report regardless of the mode, so you can simplify your own panel and still send the client everything.
In the content structure map, a section matching the article title poorly always got the same advice: consider removing it. For a short digression that is right and it stays, but for large sections we were advising to delete more than a third of the article, when the topic deserves its own piece. Across 2,218 audits such sections are common but almost always small. A section taking at least a fifth now gets a "split into own article" label with its share of the text, exports included.
Content pasted from Word, Google Docs or a CMS editor reached the audit without formatting, so the audit flagged missing structure in material that had it. Pasting now reads the formatted version and converts it to Markdown: H1-H6 headings, tables, nested lists, bold, italics, quotes, links and code blocks. Editor ballast (colours, fonts, base64 images) is stripped, and an "Undo" waits under the field.
Some phrases mean two different things in Google. For "stairs" the results split between a documentary series and stair manufacturers, and the audit treated both as competitors: hence a recommendation about episode counts and an average that included Wikipedia. We now filter out results about the other thing sharing the name. They no longer feed recommendations, averages or content comparisons, but stay in the TOP 10 marked "different topic". Second change: a keyword must contain at least two words, because a single word carries no intent. Existing audits are unaffected.
Another extension of the public API (v1.2), mostly built on feedback from integrators. You no longer have to poll us: give a webhook address when submitting and we will send the result there, including for a failed audit. Every event is signed and the verification secret is in the account response; without that check, someone could feed you a fake result. An audit no longer requires a published page: pass the text itself and get the same report at the same price. There is now a project list and a credit balance with its history, the audit list can be filtered by date range and domain, and your request limit is visible in the headers of every response.
A major extension of the public API (v1.1). You can request an audit for up to 50 URLs in one request: they go into a queue, and keyword data is fetched for the whole batch in one call, which is cheaper and faster than sending them individually. The result looks the same: the audits show up in your list, each with its own report. You can assign the whole batch to a project when submitting. A job listing recovers identifiers you lost. A report already run can be shared without a credit, with an optional expiry; revoking access is permanent. A bulk request can be cancelled, and credits for audits that never started return to your account. We also enforce API key scopes now.
Until now every fix was applied on its own and nobody looked at the finished article as a whole, which is where repeated claims or a list starting at step two show up. So we added a final check: once all changes are in, the tool reads the result as a whole, verifies that each recommendation was carried out and fixes what it finds with targeted edits, so your before/after highlights stay in place. It settles contradictions only by removing the disputed fragment, without guessing which version is true. Separately: a heading is no longer clipped in half when one section title is the beginning of another.
A broad overhaul of the "Apply to content" tab. No more duplicates: for content without Markdown headings, applying added sections next to paragraphs on the same topic and the article nearly doubled. We now give a heading to content that is already there, and skip the change when it is missing. Fewer skipped recommendations: notes about the same section reach the model together and merge into one coherent change instead of blocking each other. New: you can also apply the proposed heading title changes from the content structure. Highlighting after a manual edit now compares content rather than notation. Remember that "Regenerate" replaces manual edits.
No more rewriting the article by hand. The audit report now has a new "Apply to content" tab: you pick which recommendations (plus title/meta and heading fixes) to apply, click, and the tool surgically applies the selected changes to your content while leaving the rest untouched. You get the result in a single editor with three modes (Preview / Visual editor / Markdown), with new and changed fragments highlighted in green - you instantly see exactly what changed and which recommendation it came from (on hover). Each recommendation shows the estimated CQS gain, so it is easier to decide what is worth applying. You can still fine-tune the text by hand, save it, and copy the finished piece straight into your CMS (formatting preserved) or as Markdown. A generation costs 1 credit. You can safely close the tab while it runs - the work continues in the background and the result is waiting when you come back.
We improved the "SERP Validation" panel. First - Central Entity alignment (what your page is about vs what the search result is about) is now judged deterministically, based on semantic similarity, rather than the model's "soft" call alone. Previously the model could be too lenient and show "full alignment" even when your page's entity was a proper name or brand, while the search result was about a general concept (a role, category or topic). Now such a case gets an honest "partial", and a clearly different topic gets "low". We also refined the other signals in this section: a content intent that is close to (but not identical with) the search intent now gets "partial" instead of a harsh "low"; a page that merely links to a tool (e.g. a calculator on a subpage) is no longer mistaken for the tool page itself; and the summary text no longer contradicts the colored markers next to individual fields. Second - we tidied the layout: key data, SERP themes, PAA signals and related searches now sit in four equal columns with a consistent list style, and the "X compared/verified facts" counter is gone from the "Fact consistency" and "Consensus alignment" cards (it was almost always 1 and only confused). The panel is more compact and less spread out vertically.
We tuned the AI model's "temperature" separately for each stage of the audit, so that analytical tasks - entity and fact extraction (Knowledge Graph), intent alignment (CSI), information density, TF-IDF - run as deterministically as possible, while keeping creative freedom only where it genuinely helps (e.g. generating topic and sub-query suggestions). The result: re-auditing the same page yields much more stable dimension scores and less random fluctuation in CQS/Citability between runs. The change is entirely under the hood - it doesn't change how you use the tool or how the report looks.
The audit now checks objective, numeric facts on your page (prices, dimensions, technical specs) against the value most competitors in the search results agree on. If your number clearly departs from the consensus (e.g. a different battery capacity or a different price than most of the market), we show it in a new "Consensus alignment" card - because AI systems cross-check facts across sources, and a value that stands apart lowers the page's credibility as a source. It is a warning, not a verdict: competitors can be wrong too, so treat it as a "check whether your value is current and correct" signal. We deliberately skip fields where differences are normal - colors, sizes, variants, descriptions, and dates - and require a real consensus (at least 3 competitors with a numeric value, at least 60% agreeing). When everything matches, the card shows a green confirmation. Alongside this, the internal "Fact consistency" check (contradictions within a single page) now also confirms the positive state - "no contradictions found" - when there was actually something to compare. Both sections are warnings and do not affect the CQS or AI Citability score; they appear in the report and in exports (PDF/Markdown).
The audit now checks whether your content states two different values for the same fact - e.g. a different price in the price list than in the description, a different date in the heading than in the body, a different number for the same parameter. Such internal contradictions lower the chance AI Overview cites the page (generative systems cross-check facts and skip sources that contradict themselves). Conflicts appear in a new "Fact consistency" card in the report, as a separate recommendation, and in exports (PDF/Markdown). Conflicting dates are flagged more gently ("possibly stale"). The mechanism resists false alarms: it ignores pure format/unit differences ("10 h" = "10 hours"), multi-value attributes (color, size), and requires both values to actually appear in the text. It is a warning - it does not affect the CQS or Citability score.
The audit now recognizes e-commerce pages from the structure of the page itself (Product/Offer markup, prices, "add to cart" buttons) and recommends the right schema instead of generic ones. For a product category/listing page it recommends an ItemList of Product items (Required) plus Product with offer data (price, availability, currency) - the exact structure that lets Google show a "product list" result. For a single product it keeps Product with offer, rating, and reviews. Blogs and post listings (no prices or cart) still get editorial recommendations (CollectionPage + ItemList + Blog), so nothing changes for them. The same consistent recommendations now apply both in the audit report and in the "Schema Gaps" tool (bulk sitemap analysis) - and in the tool schema names are now localized (PL/EN) with a clear "Required" marker.
The report builder now lives in the Export tab - configuring the report layout and exporting it are in one place. In the Export tab you'll first see the export options, and below them the builder with blocks and templates. The separate "Report builder" tab is gone, and your saved templates and settings keep working unchanged.
Heavier audits (large pages, many competitors in the results) now fail with a timeout far less often - we extended the allowed analysis time and skip URLs that can't be analyzed as content (e.g. PDF files among search competitors), which also speeds up the benchmark. And if an audit still fails on our side, the credit is automatically refunded to your account - with a clear message during the audit, so you don't have to track it yourself.
Effort Score (how much structural effort went into the content) no longer applies criteria that don't fit a given page type. A table of contents is now recommended only for articles, guides, and encyclopedic content - for product pages, listings, FAQs, landing pages, and tools it is no longer required and won't lower the score. A visible publication/update date now counts for every content type except landing pages (where a visibly stale date tends to hurt rather than help). Additionally, when two independent signals - the content type from search results and the structure of the page itself - both point to a product listing or a product page, the report and exports (PDF/Markdown) flag the audit with an "e-commerce" label.
In Bulk audit you can now provide your own keyword for each URL instead of relying on automatic detection. In the list, paste pairs using the "URL ; keyword" format (one per line). For those URLs we skip automatic keyword detection and use yours - the preview marks them as "custom keyword" and you can still edit them before starting. URLs without a keyword work as before: we detect it automatically.
Every audit runs a dozen-plus independent AI analyses in parallel. Occasionally one of them momentarily fails (a timeout, a network glitch) - previously that dimension counted as 0 and unfairly lowered the Content Quality Score and AI Citability, with no indication to you. Now it works differently: (1) we automatically retry the failed analysis, (2) if it still cannot be computed we exclude it from the score and recalculate over the remaining dimensions instead of averaging in a misleading 0, and (3) the report shows a clear banner that the result is partial, suggesting you re-run the audit. The result: scores are more stable, fairer, and more comparable across audits.
We reworked how the recommendation list is built so it is more trustworthy. "Critical" now means a real problem - a dimension that genuinely scores poorly (below threshold), not just a high weight rated as high-impact by the AI; no more lists of a dozen "critical" items where everything seems equally urgent. We also removed contradictory advice (e.g. "remove section X" next to "expand section X") and duplicate recommendations coming from different dimensions. Most importantly: a quote from your content (BEFORE) is shown only when it actually appears in the article - if it cannot be verified, we hide it instead of showing a fragment that the AI may have fabricated. This applies to the in-app report, the PDF/Markdown exports, and the share link.
Bulk audit is now available to everyone. Paste a list of URLs or provide a sitemap (up to 50 pages), and the tool crawls the content, detects the keyword and infers intent (CSI) for each URL - you can still tweak keywords and uncheck pages to skip before starting. After you confirm, the audits go into a queue and run in the background (up to 3 in parallel), so you don't have to keep the tab open. The result is a folder of individual audits with a link to each page's full report. Crawling and preview are free - you pay only on confirm, 1 credit per audited URL; unused credits are refunded on cancel. An aggregate comparison report for the whole batch is coming soon.
You can now run audits via an API and fetch the result as JSON - for your own integrations and automation. Create API keys in the "API" section (button in Settings): the key is shown only once and can be revoked anytime. Usage: POST /api/v1/audits with a URL (optionally a keyword and language) returns a job id, and GET /api/v1/audits/{id} returns status and the full result (CQS, AI Citability, dimensions, recommendations, benchmark). Each API audit costs 1 credit, just like in the app. Full documentation with curl examples lives on the dedicated API page. Audits created via the API are flagged in the panel.
Effort Score now checks whether the content has a visible date - AI Search prefers fresh, maintained sources. The date is detected from four places: meta article:published_time and article:modified_time, the <time> tag, and JSON-LD (datePublished/dateModified), so it works both on pages with meta tags and those relying on structured data. The result is qualitative, not just a yes/no: below the Effort checklist (and in the PDF/Markdown exports) you see the actual dates and the content age, e.g. "Published: 2024-03-15 · Updated: 2025-11-01 · (7 mo. ago)", or a clear "No visible date". Age is computed at audit time. The analysis is fully algorithmic (0 extra AI calls) and available when auditing from a URL.
You can now audit German content - alongside Polish and English. The language is detected automatically from the crawled content, so there is nothing to configure. All AI analysis (recommendations, BEFORE/AFTER, value-add) is produced in German, and the algorithms recognize German page noise (Impressum, AGB, Datenschutz, cart, newsletter) and word inflection (stemming) so results stay accurate. The interface stays Polish and English; report labels for DE content are localized where it matters, with English shown as a universal fallback elsewhere.
We now strip far more boilerplate from crawled content that was dragging down analysis quality (TF-IDF, density, section coherence). Instead of a fixed phrase list, rules recognize whole classes of elements (PL and EN): comment forms, share buttons, cookie/GDPR banners, e-commerce sort bars and menus, cart/checkout, newsletter signups, "related / recommended" sections, rating widgets (stars), copyright footers, "follow us", ad labels, breadcrumbs, pagination, tables of contents, "no results found" and captcha. We deliberately keep author, dates and cited sources - those are E-E-A-T signals, not noise.