Geovory 有哪些新变化

过去有两个地方在 AI 调用卡住时会一直转圈。现在注册预填（根据你的网站猜测企业信息）和查询拆解工具在浏览器端和 Worker 端都有时间上限——查询拆解每次调用预算 45 秒，仅在剩余时间充足时做一次质量重试，整体不超过 90 秒的客户端上限——超时后显示明确的重试状态和预计耗时，而不是无声地转圈。

网页应用的可用性探测改为通过 Cloudflare 服务绑定进行，不再走公网，DNS 或 CDN 抖动不会再被误报为故障。

建议面板现在按 AI 模型拆分结果。不再是一张笼统的「提升可见度」卡片，而是明确告诉你哪个引擎把你漏掉了——「Perplexity 在 6 个买家问题中有 5 个跳过了你」——并列出它转而推荐的竞品，下一步动作因此针对到具体引擎。

高影响卡片最多展示三张，让面板保持待办清单的形态而不是一面墙；其余按影响力排序归入中等。

私有工作台链接现在会在边缘节点先行校验，格式错误或猜测出来的令牌会立刻返回，而不是渲染一个空白工作台。

公开接口（快速检测、查询拆解、注册）的限流现在以真实访客 IP 为键：前端通过共享代理密钥转发访客 IP，没有密钥时使用边缘节点记录的连接 IP，直接调用者无法冒用他人的配额。

评分的可信度取决于背后的回答。现在每份报告的覆盖率徽章都会统计本次扫描未能触达的通道；当 ChatGPT、Gemini、Claude 这三大主流引擎都没有返回可用回答时，评分会标记为**临时评分**并给出中性结论，而不是一个看起来很确定的字母等级。

免费快速检测也同样披露失败样本——「采样 12 条回答，4 条失败」——漂亮的分数再也不会掩盖单薄的样本量。

产品不变，品牌焕新。RankedByAI 正式更名为 Geovory——一个更短、更易记的品牌，承载同样诚实的 AI 可见度监测：真实采样回答、透明方法论、绝不编造数据。

你的账号、报告和链接均不受影响。域名将随后迁移至 geovory.com，所有现有链接都会永久重定向。

增长数据过去每周一靠手工汇总。现在由周一的定时任务自动收集——扫描量、注册量、内容与可用性探测——汇入管理后台的 Growth 版块，有异常时发送提醒邮件。

同样的数据，零手工操作，还有了一条可供逐周对比的稳定基线。

扫描结果不错的企业希望能把它展示出来。现在有了可嵌入的 AI 可见度评分徽章——由我们的 API 实时输出的 SVG——报告页新增嵌入对话框，直接复制代码即可使用。

徽章会链接回对应的报告，所以别人网站上展示的分数，永远是我们实际测得的分数。

AI 可见度带来了一堆新词汇——提及率、grounding、引用缺口、提示词采样。新上线的 /glossary 用全站六种语言定义了其中 30 个术语，每条都有通俗解释，并链接到该概念在产品中出现的位置。

博客文章和报告现在直接链接到术语表，不用再在每篇文章里重复解释同一个词。

比较 AI 可见度工具的买家总在问同一个问题——Geovory 和 X 有什么区别？新的 /alternatives 版块正面回答：八个竞品对比页，每页覆盖全站六种语言，配有诚实的功能对照表和价格背景。

如果某个竞品确实更适合某类用户，页面会直说。对比页只有经得起同时了解两个产品的人阅读，才有价值。

Geovory now looks like one product everywhere it shows up. We shipped a full logo suite — mark, wordmark, mono and inverse variants — plus social and OG image templates, a PWA manifest with app icons, and a v2 brand book that pins down how all of it is used.

The blog got the most visible upgrade: every article now renders its own dynamic OG image, so shared links carry the article's title instead of a generic banner.

Until now, publishing a blog article meant a code deploy. That coupling is gone: posts moved into a D1-backed CMS served by a worker content API, with an admin editor for writing and publishing.

The frontend renders articles via ISR, so a newly published post appears on the site within minutes — no build, no deploy, no engineer in the loop.

Day-to-day operations used to mean poking at the database directly. We built an internal admin console — overview, customers, scans and content sections — behind a Bearer-token API.

It is deliberately boring: see what is happening, find a customer, inspect a scan, edit content. Boring is what you want at 8am when something needs checking.

The fourth revision of our design system landed together with a sweep through every page. The demo card grid was rebuilt, product screenshots were recaptured at retina quality, and shadow tokens were standardized.

Alongside the new spec we fixed dozens of small alignment issues that had accumulated across the marketing site — the kind of one-pixel debts nobody files a ticket for but everybody notices.

The customer dashboard used to open with charts and leave the interpretation to you. It now leads with triage: what problems were detected this week, and what the concrete next actions are.

Every section closes with a plain-language conclusion, so the dashboard reads like an analyst's summary rather than a wall of metrics.

Our marketing pages talked about AI assistants as if every visitor lived in the US. Now the examples adapt to each of the six site languages — the assistants and search habits referenced match what buyers in that market actually use.

A small copy change, but it removes the constant translation dissonance of reading about a tool nobody around you uses.

The landing page, industry pages, blog, tools, methodology and legal pages were rebuilt in Next.js on the v3 token system: a proper typography scale, the emerald palette, card shadows, and retina product screenshots throughout.

This was the visible half of the migration off the original worker-rendered pages — same content, dramatically more coherent presentation.

geovory.com now serves the migrated pages from a Next.js app through a service-binding proxy. The cutover was designed to be reversible: a KV kill switch flips traffic back to the old renderer instantly if anything misbehaves.

Visitors saw nothing change during the switch — which was exactly the point.

Counting brand mentions is easy to get subtly wrong: 'Joe's Plumbing' and 'Joes Plumbing LLC' are the same business, and a mention buried in a caveat is not the same as a recommendation. Scans now attribute every brand mention at sentence level and normalize entity names with an LLM pass.

The result is that reports can show you the exact sentence where you — or a competitor — appeared, and variant spellings no longer split one brand's count in two.

Scans gained a third channel: live Brave web search results are fetched for each buyer question and fed to the model before it answers. This approximates how retrieval-augmented assistants actually compose recommendations — from the live web, not just training memory.

Reports now split results per channel, so you can see what models believe on their own versus what the live web tells them. The gap between the two is often the most useful finding in the report.

Scans used to run inside a single long HTTP request — workable, until a slow model response pushed it over the edge. They now run as background jobs on a Cloudflare Queue, with the browser polling for progress.

Scans are more reliable, progress is visible question by question, and a flaky connection no longer kills twenty minutes of sampling.

Knowing you were mentioned is half the story; knowing which pages AI answers cite is the other half. Reports now break down the exact URLs cited across all sampled answers, aggregated by domain, with a citation funnel view.

The deep-read insights highlight the pages your competitors win with — the comparison posts, directories and review pages that keep feeding AI answers. That list is effectively your content roadmap.

Question prompts are now auto-generated per business and industry instead of drawn from a fixed pool, and sampling scaled up to 29 question templates per scan.

A cost circuit breaker caps runaway scans, so the larger sample stays economically sane. More questions, better tailored — the mention-rate estimate got noticeably steadier.

Findings are only useful if they tell you what to do. Every detected gap now comes with a concrete page-level fix: which page to create or edit, and what it should say.

Instead of 'your FAQ coverage is weak', the report says 'add a page answering this question — here is the structure'. That difference is what makes people actually fix things.

Scans now audit your own website, not just the AI answers about you. The diagnosis checks the technical and content signals AI assistants rely on — crawler access, structured data, llms.txt, FAQ content and more — and ties each finding to the visibility gaps in your report.

It turns 'AI doesn't mention us' from a mystery into a checklist.

The Monitor and Optimize plans became free for the duration of the open beta — full weekly monitoring, prompt management and email reports, no credit card.

We want trend data and honest feedback more than early revenue. Planned post-beta pricing stays published on the pricing page so nobody is surprised later.

One scan is a snapshot; the interesting question is whether the picture is changing. Every customer now gets a private dashboard — grade, mention rate, trend chart and scan history — fed by weekly scheduled re-scans via cron.

Same questions, every week, no manual work. This is the feature the rest of the product is built around.

AI visibility looks different for a dentist than for a SaaS company, so we shipped programmatic /for/<industry> pages for twelve industries, each with JSON-LD, sitemap entries and cross-links.

Each page speaks to how buyers in that industry actually ask AI assistants for recommendations.

The entire site and all reports were localized into English, Spanish, German, French, Portuguese and Chinese. Language lives in the URL — /de/pricing, /zh/blog — with hreflang wiring throughout.

AI answers themselves stay verbatim in whatever language the model answered in: translating evidence would falsify it.

The first public version went live on Cloudflare Workers: a free AI visibility scan with a permanent report link, A–F grading, competitor extraction from real AI answers, an SEO blog, and Paddle checkout.

Rough in places, but the core promise worked on day one — ask the AI the questions your buyers ask, and show you exactly where you stand.

Shipping a fix and wondering whether it worked is a bad loop. Now you mark a recommended fix as done and re-check it instantly — the audit re-fetches the affected page and re-runs that specific check on the spot.

Trend charts show fix markers at the point each fix shipped, and weekly change summaries confirm what actually improved. Action and outcome, finally on the same chart.