The Map Wars · Part 7
You tell the assistant in your car to book "that Sichuan place from last time." It books one — a different branch with the same name, half a city away. You ask it to route to the nearest fast charger; you arrive to find the charger was removed three months ago.
Errors like these have little to do with how smart the model is. The problem sits one layer lower: AI has no stable way to refer to a "place."
In its July release this year, the Overture Maps Foundation upgraded its own positioning from open map data to "grounding AI" — and the core of that bet is something called GERS, the Global Entity Reference System. This piece takes it apart: what it actually solves, what it means for agents, and which of its promises you shouldn't rush to believe.
An Old Problem in a Newly Dangerous Setting
The same restaurant carries one ID in Amap, one place_id at Google, another at a review platform, a fourth in OpenStreetMap. Names change, coordinates drift, businesses move, close, change hands. In the machine world, "place" has never had a unified identity.
The industry's legacy fix is called conflation: fuzzy-matching on names plus coordinates. Expensive, slow, never reaching 100% accuracy — and it has to be redone for every pair of data sources. Overture has published a figure: organizations spend up to 90% of their time and resources on data preparation and integration; the cost of integrating often exceeds the cost of licensing the data. The industry has a name for the bill: the conflation tax.
For twenty years, that was the GIS engineer's pain. Now it has become the agent's fatal exposure.
The difference is action. A human misreads a map and self-corrects; an agent converts language directly into action, so a reference error becomes an action error — wrong booking, wrong routing, wrong payment. The price of hallucination has been upgraded from "saying it wrong" to "doing it wrong." Doing it wrong costs money.
What GERS Sets Out to Do
GERS reached general availability in June 2025. The design is plain: issue every spatial entity — buildings, POIs, road segments, addresses, administrative divisions — a 128-bit UUID that stays stable across data releases. Four components make the system: a monthly reference map, changelogs, a registry, and bridge files.
Bridge files are the decisive design. They are mapping tables between GERS IDs and each provider's own IDs — Meta's places map back to Facebook page IDs, OSM features map back to OSM IDs. With them, data sources no longer match pairwise against each other; each aligns to GERS once. What used to be an n-squared matching project across n sources becomes n alignments. Conflation projects measured in weeks become a database column join, measured in minutes — CARTO has demonstrated the path on Databricks, joining millions of location records directly by ID.
Once third-party data is matched to Overture entities, it becomes what the ecosystem calls GERS-enabled — joinable by ID against any other dataset in the system. This June's Places update added a field worth noticing: operating status — anonymized member signals judging whether a business is still open. A place stops being a coordinate and becomes an entity with a lifecycle.
In one sentence: DNS manages web addresses; GERS wants to manage physical ones. A household registry for the physical world.
For Agents, This Runs Four Layers Deep
The first layer is grounding. An agent's first act in any task is resolving a natural-language reference into a machine-consumable identifier — "that Sichuan place from last time" must become a definite ID before a booking API can accept it. Today that resolution happens inside each app, on private IDs, and the result cannot leave the app's ecosystem. With a public ID, references travel across systems without distortion.
The second layer is handoff. The agent economy will not be one agent doing everything; it will be relays — a booking agent, a navigation agent, a payment agent passing the baton. Every handoff must answer the same question: which place do you mean? Without a shared ID, every handoff is a translation, every translation leaks accuracy, and error rates compound with chain length. Protocols like MCP solve "how to call the tool"; entity layers like GERS solve "are we talking about the same thing." Only with both layers does the chain connect.
The third layer is memory. An agent's long-term memory has to hang on stable anchors — your regular spots, your seating preference, last time's complaint. The shop renames or relocates: if the ID persists, memory persists; if the ID changes, the agent's knowledge of you resets to zero. On unstable references, there is no such thing as real personalization.
The fourth layer is audit. Once agents book and pay on people's behalf, errors must be reconcilable: which place exactly did it mean, on which version of the data? Stable IDs make actions traceable. Whether payment and insurance — the real-money scenarios — dare to plug in agents at all rests on this.
Three Things Not to Believe Yet
Stability deserves a question mark. The June 2025 GA itself involved a reset of the ID scheme, with official mapping files required to migrate old IDs to new UUIDs. Critics this year have questioned what share of IDs get retired and reissued in monthly releases — if a meaningful fraction changes every month, the "persistence" promise collapses and the conflation tax returns in a new costume. Overture's answer is open-source matching libraries preserving IDs across versions, with feature-level conflation on the Buildings and Places themes. Whether the promise holds will be read in the next several changelogs, not in the white paper.
The standards fight is a fight over citation. Early this year the OGC has been advancing GERS toward community-standard status; at the same time, the OpenStreetMap community pushed back hard — accusations of corporate enclosure, opaque governance, and a $300,000 membership toll. On the other flank, Google will not adopt GERS — the place_id ecosystem is a limb of its own agent strategy — and Apple and Amap keep their own schemes. The registry-of-places seat will most likely not be winner-take-all but several coexisting systems, with bridges built between bridge files.
China is a parallel world. Amap, Baidu and Tencent POI IDs don't interoperate; the same conflation tax is paid in full inside the mini-program ecosystem. GERS won't enter — but the "registry for places" question will be answered in China sooner or later, and the party best positioned may not be a map company at all, but a super-app holding payment and in-store transactions. Whoever's ID gets cited by the transaction owns the registry.
Landing It in the Car
The cockpit agent is where spatial hallucination costs the most. Routing a low-battery car to a charger that doesn't exist isn't an experience blemish; it's a stranding incident. Real-time charger availability, waypoint planning, the navigate-book-pay task chain — every link presses on entity alignment.
So one concrete suggestion for teams making export-cockpit decisions: next time you select a data foundation, add one question beyond the spec sheet — what is your entity-ID system, how does it stay stable across versions, and which mainstream ecosystems does it bridge to? The question sounds premature today. In two years it will be a standard line in the procurement template.
In the last era, maps were read by human eyes; accurate and attractive was enough. In this era the map's first reader is a machine, and the first requirement shifts from attractive to citable. The registry matters more than the atlas. Whoever turns their registry into the industry default has taken a seat at the records office of the agent economy.
Sources: Overture Maps Foundation GERS documentation and June 2025 GA announcement; GERS workshop materials; OGC GERS SWG charter; Linux Foundation announcement of July 13, 2026; CARTO engineering blog.
《地图战争》系列 · 第7篇
你对车里的助手说，订上次那家川菜馆。它订了，同名的另一家分店，隔着半个城。你让它导去最近的快充桩，到了发现桩三个月前就拆了。
这类错误跟模型聪不聪明关系不大。问题出在更底下一层：AI对"地点"没有稳定的指称。
Overture Maps Foundation今年7月的通稿里，把自己的定位从开放地图数据升级成了"为AI做接地"（ground AI），核心押注就是一套叫GERS的东西——Global Entity Reference System，全局实体引用系统。这篇把它拆开讲：它到底解决什么，对Agent意味着什么，以及哪些承诺先别急着信。
一个老问题，换了个要命的场景
同一家餐厅，在高德是一个ID，在Google是一个place_id，在大众点评是另一个，在OpenStreetMap又是一个。名字会改，坐标有偏差，店会搬家、关门、换老板。机器世界里，"地点"从来没有过统一的身份。
行业的旧解法叫conflation，数据融合：拿名字加坐标做模糊匹配。又贵又慢，准确率永远到不了100%，而且每对数据源都得重做一遍。Overture给过一个数：企业最多把90%的时间和资源花在数据准备和整合上，整合的成本经常超过数据授权费本身。业内给这笔开销起了个名字，conflation tax，融合税。
过去二十年，这是GIS工程师的痛。现在它变成了Agent的命门。
区别在于行动。人看地图，看错了自己会纠正；Agent把语言直接变成行动，指称错误就是行动错误——订错、导错、付错。幻觉的代价，从"说错"升级成了"做错"。做错是要赔钱的。
GERS想干什么
2025年6月，GERS正式GA。设计很朴素：给每个空间实体——建筑、POI、路段、地址、行政区——发一个128位的UUID，跨数据版本保持稳定。整个系统四个组件：月度参考地图、变更日志、注册表、桥接文件。
桥接文件是最关键的设计。它是GERS ID和各家自有ID之间的映射表——Meta的地点数据能映射回Facebook主页ID，OSM的要素能映射回OSM ID。有了它，各数据源不需要两两互相匹配，各自对齐GERS一次就够了。n个数据源之间本来是n平方次匹配工程，现在变成n次。原来以周计的融合工程，变成一次数据库列连接，分钟级。CARTO在Databricks上跑通过这条路，几百万条位置记录直接按ID join。
第三方数据一旦匹配上Overture的实体，就成了所谓GERS-enabled的数据，可以跟生态里任何其他数据集直接按ID关联。今年6月的Places更新又加了一个字段：经营状态——用成员贡献的匿名信号判断一家店还开不开门。地点不再只是一个坐标，是一个有生命周期的实体。
一句话概括：DNS管网址，GERS想管地址。给物理世界上户口。
对Agent，这是四层意义
第一层是接地。Agent执行任务的第一步，是把自然语言里的指称解析成机器能消费的引用——"上次那家川菜馆"必须先变成一个确定的ID，订座API才接得住。今天这个解析每个App各干各的，用私有ID，解析结果出不了自家生态。有了公共ID，指称可以跨系统传递而不失真。
第二层是交接。Agent经济不会是一个Agent包打天下，而是订座Agent、导航Agent、支付Agent之间的接力。每次交接都要回答同一个问题：你说的是哪个地点。没有统一ID，每次交接就是一次翻译，每次翻译都有损耗，任务链越长，错误率越高。MCP这类协议解决的是"怎么调用工具"，GERS这类实体层解决的是"我们说的是不是同一个东西"。两层都通，链路才接得起来。
第三层是记忆。Agent的长期记忆得挂在稳定的锚点上——你常去的店、偏好的座位、上次的差评。店改了名、搬了址，ID不变，记忆不断；ID一变，Agent对你的了解就归零一次。指称不稳定的系统里，谈不上真正的个性化。
第四层是审计。Agent开始替人订、替人付之后，出了错要能对账：它当时指的到底是哪家店，依据的是哪个版本的数据。稳定ID让行动可追溯。支付和保险这些真金白银的场景敢不敢接入Agent，前提就在这里。
先别急着信的三件事
稳定性的成色要打个问号。2025年6月那次GA本身就伴随了一轮ID体系重置，旧ID迁到新UUID要靠官方发的映射文件。今年也有批评者质疑月度版本里ID退役重生的比例——如果每个月都有相当一部分ID变掉，"持久"这个承诺就塌了，融合税换个形式又回来了。Overture的说法是用开源匹配库在版本间保ID，Buildings、Places这些主题做特征级融合。承诺能不能兑现，看接下来几个版本的变更日志，不看白皮书。
标准之争的本质是引用之争。今年初OGC在推进把GERS采纳为社区标准，同一时间OSM社区强烈反弹，指责公司圈地、治理不透明、三十万美元的会员门槛。另一边，谷歌不会用GERS——place_id生态是它自己Agent战略的一部分，苹果、高德也各有体系。地点注册表这个位置，大概率不是一家通吃，而是几套体系并存，桥接文件之间再架桥接。
中国是平行世界。高德、百度、腾讯的POI ID互不相通，小程序生态里同样的融合税一分不少。GERS进不来，但"给地点上户口"这道题国内迟早有人要答——最有条件作答的未必是地图厂，可能是握着支付和到店交易的超级App。谁的ID被交易引用，谁的ID就是户口。
落到车上
座舱Agent是空间幻觉代价最高的场景。电量告急时导去一个不存在的充电桩，不是体验瑕疵，是趴窝事故。充电桩的实时可用性、途经点规划、导航到订位到支付的任务链，每一环都压在实体对齐上。
所以给做出海座舱的朋友一个具体建议：下次选数据底座，在参数表之外加一个问题——你的实体ID体系是什么，跨版本怎么保稳定，桥接到哪些主流生态。这个问题今天听着超前，两年后会是采购模板里的标配项。
上一个时代，地图的读者是人眼，做到好看、准确就够了。这个时代，地图的第一读者换成了机器，第一要件从好看变成可引用。户口本比地图册重要。谁先把户口本做成行业默认，谁就坐进了Agent经济的登记处。
数据来源：Overture Maps Foundation GERS文档与2025年6月GA公告、GERS工作坊材料、OGC GERS SWG章程、Linux Foundation 2026年7月13日公告、CARTO技术博客。